Add FP8/FP4 block-quantized support to GatherBlockQuantized contrib op

[Copilot]

Description

This PR extends GatherBlockQuantized (schema + CPU/CUDA/WebGPU kernels) to support FP8 and FP4 block-quantized data in addition to its existing integer formats.

Key semantic additions and fixes:

• Adds float8e4m3fn, float8e4m3fnuz, float8e5m2, float8e5m2fnuz, and float4e2m1 data. For these types, bits is ignored and zero_points is omitted.
• Allows block_size = 0 for FP8/FP4, meaning the full quantize_axis dimension is one block.
• Allows FP scales to broadcast on axes other than quantize_axis.
• Changes existing integer behavior as well as FP behavior: out-of-range indices now consistently produce zero output slices on CPU, CUDA, and WebGPU. This deliberately differs from ONNX Gather; it avoids provider-dependent behavior and preserves asynchronous CUDA execution without stream synchronization. Op-owner sign-off is requested for this contract change.
• Aligns CUDA defaults with the schema (block_size = 128, quantize_axis = 1) and validates axes, ranks, scale shapes, empty outputs, launch dimensions, and integer bits values.
• Validates block sizes and integer bits during graph resolution; bits remains intentionally ignored for FP8/FP4.
• Makes the WebGPU FP8/FP4 path reachable and tests constant FP tables across supported providers.

Tests cover CPU/CUDA/WebGPU behavior, defaults, BFloat16 output, packed FP4 row boundaries, negative axes and indices, invalid indices, empty outputs, and invalid attributes. Generated contrib operator/kernel documentation is included.

Motivation and Context

Qwen3.8-Flash / Qwen4-Exp requires an n-gram embedding gather over a block-quantized FP8/FP4 table. Extending GatherBlockQuantized keeps a single Gather-with-quantization op while adding the dtype and broadcast-scale support required by that table.