fix: allow restart-analysis replay in assert_cache_hit mode

[cennn]

Problem

When assert_cache_hit=True (bake-verify mode) and a subgraph has restart_analysis_count > 0, CompilerManager.load() returns None due to restart-analysis skip. compile() then unconditionally raises RuntimeError('cache miss'), even though the cache entry exists.

This causes verify_compile_cache.sh to fail for any model whose compilation triggers RestartAnalysis (e.g. via TensorifyScalarRestartAnalysis from the tensorify_python_scalars pass).

Fix

Change the assert condition from:

if self.compile_config.assert_cache_hit:
    raise RuntimeError(...)

to:

if self.compile_config.assert_cache_hit and cache_entry not in self.cache:
    raise RuntimeError(...)

When the cache entry exists but load() returned None (restart skip), fall through to the normal compile path. standalone_compile will trigger the same TensorifyScalarRestartAnalysis as bake, dynamo re-traces with the correct graph shape, and on retry load() succeeds.

Test

test_assert_cache_hit_restart_analysis.py — two-process integration test:

• Process 1 (bake): warms cache, encounters RestartAnalysis → restart_analysis_count=1
• Process 2 (verify): runs with MAGI_COMPILE_ASSERT_CACHE_HIT=1 → must succeed

Athena PR: https://github.com/world-sim-dev/athena/pull/1068