Make in-memory cache misses single-flight

[adenzhou1350]

Summary

• synchronize in-memory cache metadata and make same-key misses single-flight
• keep factories for different keys concurrent
• share a failed factory result with current waiters while allowing later retry
• release the Python GIL across the complete Runtime single-flight operation
  so a waiter cannot prevent the producer from publishing its result

Testing

• python3 tests/test_memory_cache.py: PASS
• 25 consecutive repetitions of the focused test: 25/25 PASS
• real-GPU same-key CUDA regression: 10/10 candidate-tree repetitions plus
  one exact-commit archive run PASS
• syntax-only translation unit including deep_jit/runtime/runtime.hpp: PASS
• python -m py_compile tests/test_cuda.py: PASS
• git diff --check: PASS

The focused CPU regression covers 16-thread same-key deduplication, different
keys remaining concurrent, shared failure, and successful retry. The CUDA
suite also gains a bounded isolated-process regression in which two Python
threads compile the same key and exactly one NVCC wrapper must run.

The CUDA regression ran on an NVIDIA GeForce RTX 4060 Laptop GPU (compute
capability 8.9, driver 592.82) in the official CUDA 13.0.2 Ubuntu 24.04 image
with PyTorch 2.12.1+cu130. Every run observed one NVCC invocation and one shared
non-null kernel address for two same-key Python threads, with no GIL deadlock.

The broader CUDA suite passed the extension/header/lazy-init/fork/GIL/cache
publication phases, then stopped at the existing TMA driver test with
CUDA_ERROR_NOT_SUPPORTED on pre-Hopper Ada hardware; it is not presented as
a full-suite PASS.

Scope

The thread-safety guarantee is limited to callers using
MemCache::get_or_create. Direct concurrent access to the public compatibility
container is unsupported. This change does not claim that the entire Runtime,
Parser, disk cache, compiler backend, or device loader is thread-safe.