Add get_npu_cache_dir() to expose NPU binary cache path

[erwei-xilinx]

Summary

• Add get_npu_cache_dir(compiled_kernel) utility function that returns the path to the NPU binary cache directory containing aie.xclbin/aie.elf, insts.bin, and __npu_dispatch.so
• Add on_cache_resolved callback mechanism in compile_module() so NPULauncher captures the cache directory on first kernel launch
• Only amd_triton_npu/backend/driver.py is modified

Usage

from triton.backends.amd_triton_npu.driver import get_npu_cache_dir

compiled_kernel = my_kernel[grid](a, b, c, N, BLOCK_SIZE_N=1024)
npu_cache = get_npu_cache_dir(compiled_kernel)
# npu_cache == "/home/user/.triton/cache/XXXXX/"
# Contains: __npu_dispatch.so, aie.xclbin (or aie.elf), insts.bin

Closes #41

Test plan

• Basic API tests pass (None input, missing _run, wrong _run type raises TypeError)
• Triton compile stage (TTIR → ttsharedir) works, returns None before launch
• Hardware test on npu1 (Phoenix): get_npu_cache_dir returns valid cache directory with aie.xclbin, insts.bin, __npu_dispatch.so
• Verified on both cache-miss (first compile) and cache-hit paths
• All npu1 examples pass: relu, axpy, sigmoid, vec-add, matmul_bf16

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR exposes the NPU backend’s binary cache directory for a given compiled Triton kernel by capturing the cache directory during first launch and providing a helper to retrieve it from the compiled kernel object.

Changes:

• Extend compile_module() with an on_cache_resolved callback invoked with the NPU cache directory.
• Store the resolved cache directory on NPULauncher instances at launch time.
• Add get_npu_cache_dir(compiled_kernel) utility to return the captured cache directory (or None before first launch).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[ypapadop-amd]

LGTM