Refactor channel position to x/y/size/headstage to match CereLink #184

[cboulay]

This branch updates ezmsg-blackrock to track the channel-metadata rework in CerebusOSS/CereLink#184. Upstream repurposed cbPKT_CHANINFO.position[] to carry true electrode geometry (x, y, size, headstage_id) in micrometers, changed pycbsdk.cmp.parse_cmp to key entries by device (bank, term) with flat fields and verbatim labels, and moved bank/electrode out of position into the chaninfo BANK/TERM fields. Our CereLink*Producer device path and our offline ChannelMapProcessor both consumed the old layout, so both needed updating. While here, the per-channel ch axis dtype was reworked: x/y/size are now int32 (matching the library's native storage), a headstage field was added, and the unused device field was removed.

Motivation

In the old layout, position[] held (col, row, bank_letter_index, electrode) and parse_cmp returned a dict keyed by ordinal chan_id with an entry.position 4-tuple and hs{N}--prefixed labels. CereLink #184 frees the position slots to hold real geometry and matches CMP rows to live channels by the device's own (bank, term) instead of a fragile ordinal assumption. Without these changes our channel x/y would be misread (the slots now mean something different), bank/elec would be garbage, and ChannelMapProcessor would crash outright because entry.position no longer exists and the parser keys are now (bank, term) tuples rather than integers.

What changed

Device path — src/ezmsg/blackrock/cerelink.py

• _cache_channel_metadata now reads bank/term from chaninfo via get_channels_field(ChannelType.FRONTEND, ChanInfoField.BANK/TERM) (these fields predate #184 and exist in both pycbsdk versions) and keeps the full position tuple (x, y, size, headstage_id). The cache value is now ch_id -> (x, y, size, headstage, bank_num, term).
• _build_ch_info populates the new size and headstage fields, takes x/y from position[0:2] (now micrometers), and derives bank/elec from the cached chaninfo bank_num/term rather than the old position[2:4].

CMP overlay path — src/ezmsg/blackrock/channel_map.py

• The overlay loop iterates the new (bank, term) keys and computes the channel index directly as (bank - 1) * 32 + (term - 1); start_chan is already folded into bank upstream via its // 32 bank offset, so no separate ordinal handling is needed. It writes entry.x/entry.y/entry.size (cast to int), entry.headstage, the verbatim entry.label, and derives bank/elec from the key.
• _fill_auto_grid now scales its grid step to the CMP's detected electrode pitch (new _cmp_pitch helper, ≈400 µm for a default Utah array, falling back to 1 when there is no CMP) so auto-laid channels share the CMP's micrometer scale instead of clumping near the origin. Auto-grid channels get size = step and headstage = 0.
• The base-layer carry-through block was removed: it existed only to copy the device field forward, which no longer exists, and every dtype field is now set by the label/overlay/auto-grid passes.

CHANNEL_DTYPE

• x, y changed from float32 to int32, matching how the library stores position. A new int32 size field (electrode size in µm, 0 = unspecified) and int32 headstage field (1-based headstage id, 0 = none/auto) were added. The unused device field was removed — stream provenance will instead be folded in naturally during concat from each AxisArray's .attrs, rather than being a per-channel column nothing populated.

Behavior changes worth calling out

• Labels are now verbatim. hs_id no longer prefixes labels with hs{N}-; it populates the headstage field instead. Two channels on different headstages can now carry identical label strings, distinguished by headstage/bank/elec. Any downstream code keying channels purely by label string should be aware.
• Coordinates are micrometers, as integers. x/y/size are now real geometry in µm (the manufacturer index grids are scaled by the 400 µm Utah-array pitch upstream), not column/row indices.
• start_chan is quantized to whole banks on the CMP path. Because #184 applies start_chan as a // 32 bank offset rather than a per-row ordinal, values that are not of the form 1, 33, 65, 129, … snap to a bank boundary. This differs from the old arbitrary-offset behavior.
• CHANNEL_DTYPE field set changed (added size, headstage; removed device; x/y now int). Consumers that read these fields by name are unaffected by the additions/removal; any code that assumed float x/y or the presence of device will need a look.

Dependency note

These changes require a pycbsdk build that includes CerebusOSS/CereLink#184, which is not yet released to PyPI. During development the working tree points pyproject.toml's [tool.uv.sources] at a local editable checkout of the patched pycbsdk; that local-path source is machine-specific and is not intended to merge as-is. Before merging, the pycbsdk floor in [project.dependencies] should be bumped to the released version that contains #184 and the local [tool.uv.sources] entry dropped. Note also that exercising the live device path against hardware additionally requires a rebuilt libcbsdk native library from the #184 C++ source; the pure-Python parse_cmp (CMP overlay path) needs only the Python package.

TODO

• bump pycbsdk dependency