fix(cutile-rs): keep every launch-site specialization

[rdeaton]

Summary

A generated launcher's LaunchSite held only its most recent resolution, so a kernel launched alternately at two specializations (the same kernel at two shapes in one model forward, or a scalar whose divisibility changes) missed on every launch and resolved again through key construction and the kernel cache. This PR adds a larger

In the host_overhead bench on my machine(Ryzen 9 9900X3D, RTX 5080) that miss path, launch_site_miss, goes from 3.84 to 1.72 µs a launch; launch with one specialization is unchanged (1.65 / 1.66 µs).

I originally had these patches on cutile 0.3 where they mattered a bit more, but NVlabs/cutile-rs#307 cut the overhead down a lot. Still, this change makes a substantial performance difference for launch-bound workloads. My workload is the batched forward pass of an ML model for a turn-based strategy game. Each forward runs the same GEMM, layer-norm and attention kernels at several shapes, and batches are padded to a ladder of token counts, so a single-entry launch site misses on almost every launch. After 0.4, I measure a 5-10% more forward passes, about a total of 78us of host time per forward, though it was as high as > 20% before I invested in doing a lot more kernel fusion.

Changes

• The site now keeps one entry per probe it resolved. A lookup compares against the most recently stored entry first to keep the common path fast, then looks the probe up by its FxHash.
• I've new-typed the CacheEpoch to prevent inadvertent misuse
• LaunchSite, SiteResolution and the new CacheEpoch are now #[doc(hidden)] launcher support for generated code
• The tests evict only their own kernel, or bump the epoch through a new test-only hook that leaves the kernel cache alone, so they don't unload modules other tests in the binary are using. There's probably some follow-up work to clean up how tests interact with a shared cache like this, but I left it out of scope here.

Testing

• fmt and clippy;
• run_cpu_tests.sh;
• [x]the full run_gpu_tests.sh (750/750 on each branch);
• [x]Ran the host_overhead benchmarks, verified that changes to the existing launch are within noise on my hardware (Ryzen 9 9900X3D, RTX 5080).

Checklist

• All commits signed off (git commit -s)
• SPDX headers on new source files

---

Questions about the review? Ping us in #contributors on Discord.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.