[WIP] Enable MultiCastTranspose for expert weights

[sudhu2k]

Description

This PR optimizes FP8 weight quantization for GroupedLinear by replacing the per-weight quantize loop with a single fused multi_cast_transpose kernel for delayed-scaling FP8. Previously each weight in a group was cast and transposed with its own quantize call; now the whole group is processed in one fused call, reducing kernel-launch overhead and unnecessary tensor reallocations during FP8 operations.

To make this work efficiently with the existing weight-caching mechanism, multi_tensor_quantize now accepts an optional outputs argument so the fused kernel can write directly into cached workspace buffers in place rather than allocating fresh tensors on every update.

Fixes #16929

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:

• Added TransformerEngineBaseModule.get_multi_weight_workspace, a group-aware analog of get_weight_workspace that casts and transposes a whole group of weights with a single fused multi_tensor_quantize kernel for delayed-scaling FP8, with caching support. Falls back to the per-tensor get_weight_workspace path when fusion is not applicable (other recipes, rowwise-only usage, already-quantized weights, or CUDA-graph weight caching).
• Updated _GroupedLinear to use get_multi_weight_workspace instead of looping over get_weight_workspace per GEMM.
• Extended the multi_tensor_quantize C++ extension (extensions.h, cast.cpp, pybind.cpp) with an optional outputs parameter, enabling in-place quantization into pre-allocated/cached workspace buffers and avoiding redundant tensor allocations.
• Refactored the workspace allocation/caching logic to handle cache hits and misses through a unified fused path: a cache miss allocates fresh buffers, while a cache hit with update_workspace reuses cached buffers via the new outputs argument. Quantizers are temporarily forced to internal=False so cached workspaces survive prepare_for_saving.

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes