Refactor cmp

[cboulay]

Summary

Real-world .cmp files break two assumptions our previous loader made: rows are not in (bank, electrode) order, and one device may carry multiple headstages whose label namespaces (elec1-1…elec1-128) collide. This PR rewrites the loader around those facts.

• bank_offset is gone. loadChannelMap now takes (filepath, start_chan=1, hs_id=0). The parser sorts rows by (bank, electrode) and assigns chan_id = start_chan + sorted_index, so each CMP describes one headstage applied to a contiguous channel range.
• Labels are namespaced and pushed to the device. When hs_id != 0, every label is prefixed "hs{hs_id}-". The loader sends CHANSETLABEL per channel so the device persists the prefixed labels in chaninfo. hs_id=0 skips the prefix (sensible for single-headstage rigs).
• Position overlay is keyed by chan_id. Both the CHANREP receive path and the post-load batch re-apply look up by absolute channel number; positions are re-applied on every device-echoed CHANREP so a device refresh can't clobber locally-supplied geometry.
• Multi-headstage works: call loadChannelMap once per headstage with disjoint start_chan ranges (e.g. 1, 129, …). Subsequent calls merge into the overlay.

API changes (breaking)

	Before	After
C++	SdkSession::loadChannelMap(path, bank_offset=0)	SdkSession::loadChannelMap(path, start_chan=1, hs_id=0)
C	cbsdk_session_load_channel_map(s, path, bank_offset)	cbsdk_session_load_channel_map(s, path, start_chan, hs_id)
Py	Session.load_channel_map(path, bank_offset=0)	Session.load_channel_map(path, start_chan=1, hs_id=0)

New: pycbsdk.cmp

• parse_cmp(path, start_chan=1, hs_id=0) returns dict[chan_id, CmpEntry] mirroring the C++ parser.
• CLI: python -m pycbsdk.cmp head1.cmp:1:1 head2.cmp:129:2 --device NSP applies one or more FILE:START_CHAN:HS_ID specs to a live device, or --dump prints the parsed entries without connecting.

New test fixture

tests/128ChannelManufacturerMapping.cmp — sanitized 128-channel sample from a real manufacturer CMP (description scrubbed of array IDs and serial). The integration tests now point at this file so they exercise out-of-order rows.

Test plan

• C++ unit (cmp_parser_tests): 6 tests pass — sort order, start_chan offset, hs_id=0 no-prefix path, hs_id != 0 prefix path, missing-file error.
• Python unit (tests/test_cmp.py): 12 tests pass — same coverage on the Python parser plus CLI spec parsing and --dump output.
• pycbsdk integration (TestCMP, 7 tests, against nPlayServer): smoke load, error path, positions match parser, labels round-trip through device after sync(), hs_id change rewrites only the prefix, two-headstage start_chan layouts agree, CHANREP-echo doesn't erase the overlay.
• CI: GitHub Actions runs the macOS + Linux integration jobs against nPlayServer playing back dnss256 with the new manufacturer fixture as NPLAY_CMP_FILE.

Migration notes

• Single-headstage code: replace load_channel_map(path) (which used to apply bank_offset=0) with the same call — no change at the call site, but device labels are now untouched (no prefix when hs_id=0).
• Multi-port / multi-headstage code: replace load_channel_map(path, bank_offset=4*N) with load_channel_map(path, start_chan=1+N*chans_per_hs, hs_id=N+1).