[TLE] [ENFLAME] support TLE features

[baoqiliu]

[ENFLAME] Support TLE features for GCU backend
Enable TLE (Triton Language Extension) framework support on GCU backend,
including TLE-Raw MLIR EDSL, GCU Dialect EDSL (TOPS), compiler pipeline
integration, GCU300/GCU400C++ backend upgrades, and comprehensive
test coverage.

TLE-Lite现状

Python DSL	状态
tle.load(is_async=True)	✅已支持
tle.cumsum	✅已支持
tle.extract_tile	✅已支持
tle.insert_tile	✅已支持
tle.device_mesh	❌未支持
tle.shardingtle.S tle.Ptle.B	❌未支持
tle.make_sharded_tensortle.ShardedTensor	❌未支持
tle.shard_id	❌未支持
tle.distributed_barrier	❌未支持
tle.remote	❌未支持
tle.reshard	❌未支持
tle.distributed_dot	❌未支持

TLE-Struct (GPU) 现状

Python DSL	状态
tle.gpu.alloc	✅已支持
tle.gpu.local_ptr	✅已支持
tle.gpu.copy (TMA)	✅已支持
tle.gpu.memory_space	✅已支持
tle.gpu.pipeline	❌未支持

Test Result:
====================================== 4708 passed, 3482 skipped, 58 warnings in 2920.80s (0:48:40) =======================================

1. TLE Framework & GCU Dialect EDSL

1.1 TLE-Raw TOPS Runtime (new module: tle/raw/tops/) and Pass Support

- Add gcu_dialect.py: GCU MLIR dialect EDSL with native !gcu.ptr<T>
  type support, gcu.ptr2memref, vector.maskedload patterns
- Add mlir_runtime.py: MLIR-based runtime for GCU dialect IR execution
- Add runtime.py: TOPS C++ kernel integration runtime
- Extend codegen.py and runtime.py to support GCU backend dispatch
- Add gcu300.py, gcu400.py, mlir.py: typed Python wrappers for
  C++ MLIR passes via libtriton_gcu{300,400}_core
- Integrate TLE-raw pre-passes (TleToTritonGCU, ConvertArgToMemDesc,
  RemoveRedundantCopy, DSLRegionInline) with has_tle_raw detection
- Add gcu_intrinsics.py for GCU intrinsic function mapping
  Support IntrinsicsGCU/IntrinsicsTCLE data files for GCU400

1.2 TLE-Struct/Lite Core Pass Support

- Support TleToTritonGCU.cpp: TLE-to-TritonGCU conversion
  Phase 1 - Barrier Insertion
  Phase 2 - Lower TLE ops to TritonGCU ops (tle.local_pointers, ttg.tma_copy, tle.extract_ptr)

2. GCU Backend

2.1 Conversion Passes (Optimized)

- Rewrite ReduceOpToGCU.cpp: improved reduce lowering
- Rewrite ScanOpToGCU.cpp: expanded scan support
- Enhance TTSmemLowerToGCU.cpp: shared memory lowering
- Rework TritonToGCU.cpp: main conversion pipeline
- Add TritonGCULocalMemOptimize.cpp: local memory optimization pass
- Add ReduceScanCommon: shared reduce/scan utilities
- Add Vectorize.cpp/h: vectorization support
- Add TritonToTritonGPU conversion for GCU400
- Enhance PtrAnalysis, MaskAnalysis, FirstLastUserAnalysis

2.2 Python Bindings (new)

- Add triton_gcu400_core.cpp/h: C++ pass registration module
- Add triton_gcu400_module.cpp: Python pybind11 bindings
- Export pass functions via triton_gcu400_core.map

3. Tests

3.1 TLE-Raw Tests (Enflame, 13 new files)

- MLIR EDSL: vector-add, softmax, gcu_opt_lowering, ptr2memref
- TOPS EDSL: vector-add, softmax, matrix-multiplication, dialect IR

3.2 TLE Unit Tests (Enflame, 4 new files, 7 files updated)

- test_tle_raw_tops, test_extract/insert_tile_static_index
- DeepSeek V32 sparse-MLA tutorial
- Update test_tle_cumsum, test_tle_gpu_local_ptr for GCU compatibility
- Update tutorials (01-fft, 03-topk, deepseek_v32/01-topk_selector)