feat(quantization): add fused amax+scale FP8 tensorwise quantize

[jasainio]

Fuse the abs-max reduction and scale computation into a single final-round kernel, reducing total kernel launches from 3 to 2 for the tensorwise FP8 quantization path.

• Add reduce_amax_final_scale_kernel that computes scale = fp8_max / amax and scale_inv = amax / fp8_max directly in the reduction epilogue
• Add reduce_amax_and_compute_scale dispatcher with multi-round support for large tensors (reuses existing reduce_row_kernel for intermediate rounds)
• Register quantize_fp8_tensorwise_fused op in PyTorch bindings (CUDA + Meta)
• Expose quantize_fp8_fused() public Python API with fallback for non-tensorwise granularities
• Add tests for fused path: tensorwise + amax correctness (66 cases)
• Add benchmark script comparing original 3-kernel vs fused 2-kernel latency

Description

Fuse the abs-max reduction and scale computation into a single HIP kernel for the tensorwise FP8 quantization path. The original path required 3 kernel launches (reduce_row -> compute_scale_from_amax -> quantize_tensorwise_impl). The fused path eliminates the standalone scale computation by folding it into the final reduction round, reducing the total to 2 kernel launches (reduce_amax_and_compute_scale -> quantize_tensorwise_impl). For large tensors that require multi-round reduction, intermediate rounds reuse the existing reduce_row_kernel, and only the final round uses the fused kernel. A workspace size debug assertion and ping/pong buffer aliasing safety are documented in the kernel code.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• Add reduce_amax_final_scale_kernel in csrc/kernels/quantization/quantization.cu that fuses abs-max reduction with scale/scale_inv computation in the epilogue
• Add reduce_amax_and_compute_scale host dispatcher with single-tile fast path and multi-round support using ping/pong workspace buffers
• Add workspace size debug assertion and ping/pong aliasing safety comment for multi-round reduction
• Declare reduce_amax_and_compute_scale template in csrc/include/primus_turbo/quantization.h
• Add quantize_fp8_tensorwise_fused C++ implementation in csrc/pytorch/quantization/quantization.cpp with workspace allocation
• Register quantize_fp8_tensorwise_fused op in csrc/pytorch/bindings_pytorch.cpp (CUDA + Meta) and csrc/pytorch/extensions.h
• Add quantize_fp8_tensorwise_fused_impl Python wrapper in primus_turbo/pytorch/kernels/quantization/quantization_impl.py
• Expose quantize_fp8_fused() public API in primus_turbo/pytorch/ops/quantization.py with fallback to quantize_fp8 for non-tensorwise granularities
• Add test_quantize_fp8_tensorwise_fused (36 cases: 3 dtypes x 2 FP8 types x 3 numel sizes x 2 torch_compile modes) and test_quantize_fp8_tensorwise_fused_amax_correctness (30 cases: partial tile and spike-position regression)