fix(ci): use stable cache key for CUDA flash-attn builds

[starpit]

Summary

• Remove cargo registry cache step (JS actions run on host in container jobs, cannot access the container ~/.cargo)
• Use stable cache key based on features + compute cap instead of Cargo.lock hash, so the cache actually hits across releases
• On exact key match, actions/cache skips the expensive Post save step entirely (~20GB upload avoided)
• Cargo fingerprinting still recompiles changed deps regardless of cache contents

Test plan

• Trigger a release build — first run will miss and save (one-time cost)
• Trigger a second release build — should hit cache, Post save should be skipped
• Verify flash-attn binaries are correct (deps rebuilt as needed by Cargo fingerprinting)

Generated with Claude Code