add memory preflight check

[mike-ferguson]

Add preallocate_memory hook and score_benchmark helper to Benchmark

Problem

Neural benchmarks can run for 6-7 hours and crash with an OOM error at the very end
when trying to allocate the full activation array. There was no way to detect this early.

Changes

brainscore_core/benchmarks/__init__.py

• Added Benchmark.preallocate_memory(candidate) — a no-op by default. Domain-specific
  subclasses (e.g. NeuralBenchmark in brainscore_vision) can override this to run a
  pre-flight memory check before scoring begins.
• Added score_benchmark(benchmark, candidate) — the recommended call site instead of
  benchmark(candidate) directly. Calls preallocate_memory first, then scores.

tests/test_plugin_management/test_benchmarks.py — new unit test file (13 tests)
covering the no-op default, call ordering, MemoryError propagation, and env var bypass.

Design notes

• The default preallocate_memory is a no-op, so all existing benchmarks and plugins
  are completely unaffected.
• score_benchmark is the hook point — domain libraries call this instead of
  benchmark(candidate) and get the check for free.
• The actual probe logic (running 1 stimulus through the model, estimating GB) lives in
  brainscore_vision, not here. Core only defines the interface.
• Language benchmarks can adopt the same pattern by overriding preallocate_memory when
  ready — no further core changes needed.

Companion PR (vision)

brainscore_vision will add the probe implementation in benchmark_helpers/memory.py and
wires NeuralBenchmark.preallocate_memory to call it. This core PR must be merged and
released before the vision PR can be merged.

[KartikP]

LGTM! Lots of tests, but comprehensive