Keep every placement a broken pool would have lost, and drop the pin cliff

[fpedd]

A worker dying abruptly breaks the whole executor, so every future still in
flight fails with it: allocate_parallel was catching that per variant and
finding nothing left, turning one OOM kill into "Every allocator variant
failed". The variants a broken pool stranded never ran at all, so they are
retried one pool each -- never in the calling process, where whatever killed
the worker would kill the caller, and never sharing a pool, where the same
variant would strand them again.

The thread ceiling lives in native process-global state, which a forked
worker inherits and a spawned one does not, so the cap was silently absent
inside every pool worker off Linux. Pools hand it down explicitly now, split
by the worker count: the ceiling covers the product, not each level of it.

stack_around_pins fell off its linear path the moment a single pin
existed, rescanning every occupied range per item and inserting into it:
20k allocations went from 9ms to 14.9s, and the naive allocator is the
linear baseline. Claims only ever shrink a free range from below and never
split one, so the range count is fixed at what the pins leave and a max tree
over the capacities answers each first fit in O(log n).

The adjacency ceiling counts 4 bytes per directed edge, but the placers took
their rows through ConflictIndices, widening every one to size_t behind a
per-row vector. They walk the CSR rows in place now, so the accounting is
the whole cost rather than a stage of it, and the type is gone.

TwoPlusTwoSource truncated its last group below four allocations, handing
back a linearizable instance from a source that promises the opposite; it
refuses the count instead. Validating a Memory directly no longer names it
twice, and available_cores reports the benchmark environment's core count
rather than being unused beside os.cpu_count.