Add mitigations for intermittent test failures with CUDA OOM - #2484

Addresses #2381.

Test-suite hygiene for the address-space exhaustion described in this comment. Two of the three contributing factors are in the tests themselves, and this PR takes both.

Deterministic release at test boundaries. _init_cuda_context now collects and
drains before dropping the context. A pool's virtual address reservation is not
returned until the pool is destroyed and the stream-ordered frees of its outstanding
allocations retire; previously neither was forced between tests, so reservations from
earlier tests stayed live while later tests created their own.

Right-sized pools. Seven pool constructions passed options without setting
max_size, so each reserved a system-dependent window scaling with device memory while
holding at most 1 KiB:

• test_memory_peer_access.py — four device pools, including the isolated_dmr_x2
  fixture (once per test across seven tests) and a comprehension that reserved one
  window per device. All hold at most NBYTES = 1024.
• test_memory.py — three pinned pools in the numa_id tests, which only read
  mr.numa_id and never allocate.
  All now pass the suite-standard POOL_SIZE = 2097152, newly introduced to
  test_memory_peer_access.py with a comment explaining why a peer-access test caps a
  pool it barely uses.

Measured impact

On a 23.9 GiB device, 20 uncapped pools exhaust the process address space, versus
125 pools capped at 4 GiB — an uncapped pool costs roughly 1x device memory in
reservation, a 2 MiB cap costs 2 MiB.

Deliberately not included

Managed pools cannot be right-sized: cuMemPoolCreate documents that
CUmemPoolProps.maxSize must be zero for managed pools, so
ManagedMemoryResourceOptions correctly has no max_size option.

Checklist

• New or existing tests cover these changes.
• The documentation is up to date with these changes.