Parallel kernels, one accounting layer, one reverse index

[alexy]

Step 3 of the order in codex-to-codex.md (08:15Z): the rayon layer, the two
defects the review reproduced, and the merge of the two parallel layers into one.

What is here

The accounting layer. WorkMeter gives one worker a batched share of the
work budget, so N workers do not retry a compare-exchange against one cache line
once per visited entry. Grants are reclaimable: a meter that cannot be admitted
returns its own idle remainder and takes back every other live meter's, so
whether work fits a budget never depends on how many meters exist.
reserve_for_workers admits per-worker scratch once, before a region, keeping
the memory mutex off the per-element path. ExecutionContext::with_concurrency
is how an embedder says how many threads a query may add; an execution that says
nothing runs exactly the code it ran before.

Parallel kernels: degree, pagerank (weighted and unweighted),
bfs, multiSourceBfs, wcc, projectionStats. Left sequential on purpose:
dfs and topologicalSort, whose emitted order is observable; scc and
dijkstra, which are the oracles their parallel replacements will be tested
against.

One parallel module. The union of both branches': the catalog's
balanced_ranges, map_ranges, ordered_blocks, for_chunks and sort_total,
with this branch's workers, pool and meters. width() survives only until the
catalog kernels move onto workers().

One reverse index. The projection had a reachability-only reverse index and a
transpose with weights; now it has the transpose alone, and edge slots are
omitted from it because no caller reads them — about a gigabyte of arcs on a
directed graph of com-Orkut's size that SCC and PageRank would have paid to
ignore.

The rules the kernels keep

• Threads come from the execution, never from the machine: nothing reads the CPU
  count, because these kernels also run inside a server with its own pool.
• Partitions follow the work for disjoint writes, and are fixed for float
  reductions, because regrouping the terms of a float sum changes its low bits.
• Results are combined in index order, so an answer is bit-for-bit the same at
  one worker and at sixteen.
• The sequential path stays, below a measured floor and whenever the parallel
  feature is off, and it is the differential oracle in tests.

Tests

grust-algorithms/tests/parallel.rs compares every parallel kernel with its
sequential path on a 120,000-node graph and with itself at 2 and 16 workers;
grust-procedures/tests/contracts/parallel.rs holds the budget contracts at
sixteen threads. Three of those tests arrived from the review as failing
reproductions, cherry-picked unchanged in 70b7003: an idle meter's grant
refusing work that fits, the same under skew at sixteen threads, and PageRank's
scores moving with the worker count. Each fails without its fix on this box.

Numbers

Warm, sixteen workers, against the sequential kernels, on quegee (16 vCPU over 8
physical cores). examples/scaling prints what each kernel produced beside every
timing, and those values agree at every worker count.

graph	kernel	algorithm alone, 1 worker	at 16 workers
roadNet-CA	pagerank	4.0x	29.3x
roadNet-CA	pagerank, weighted	2.5x	19.0x
roadNet-CA	wcc	3.0x	14.5x
roadNet-CA	bfs	1.6x	2.6x
com-Orkut	pagerank	2.9x	22.0x
com-Orkut	wcc	3.8x	28.0x
com-Orkut	bfs	1.4x	6.7x

Three of these are better algorithms as well as threaded ones, which is why the
first column exists: reporting only the total would credit threads with work the
algorithm did. Correctness cost what it cost — the fixed reduction chunk took
PageRank on roadNet-CA from 34.4x to 29.3x — and codex-to-codex.md has the
before-and-after table.

Not in this PR

Moving the catalog kernels onto WorkMeter and workers() and retiring Meter
and width() is step 5, and the catalog agent's. Parallelising the projection
build and the transpose is step 6.