feat(moore): call MATE flash attention natively

[voltjia]

Summary

• Replace the per-call Python/MATE bridge from feat(moore): support flash-attn via MooreThreads/mate v0.2.5 #983 with a TVM-FFI native execution path.
• Use MATE only once per operator to resolve/compile the selected FMHA and combine modules, then cache their loaded DSOs and __tvm_ffi_* entry points.
• Write FlashAttention output and LSE directly into InfiniOps-owned tensors, removing provider allocations and D2D result copies.
• Route the Moore MATE glue through the existing system-C++ Torch object path so the vendor MUSA compiler does not parse ATen headers.

Validation

• Moore minimal linked build.
• Existing FlashAttention varlen/KV-cache correctness suites.
• Repeat-call check confirming Python is bootstrap-only.
• Non-default stream and split-KV combine coverage.

This is a draft while Moore hardware validation is in progress.