fix(integrations): forward distributed sharding options

[jfischoff]

Summary

FlashPack's core loader already supports both rank-0 read+broadcast and one-shard-per-rank read+replicate for arbitrary packed dtypes. The high-level integrations did not expose that contract consistently:

• Diffusers pipelines silently dropped distributed_sharded and distributed_shard_strategy, so callers got rank-0 broadcast even when they requested sharded reads.
• Transformers and Diffusers model loaders happened to pass the options through **kwargs, but did not declare or explicitly forward them.

This change makes both options first-class across all three high-level paths and documents that distributed loading works for ordinary BF16/FP32 packs as well as quantized packs.

Changes

• Forward sharded-load controls through FlashPackDiffusionPipeline and load_sub_model_flashpack.
• Declare and explicitly forward the controls in both model integrations.
• Add dependency-free AST guardrails that fail if any high-level layer drops either option again.
• Add a high-level distributed loading example to the README.

Test plan

• uv run --extra dev pytest -q -m 'not network' — 244 passed, 4 skipped, 4 deselected.
• .venv/bin/pre-commit run --all-files — all hooks passed.
• Full default test selection also passed 246 tests; its only two failures were pre-existing network integration tests importing optional diffusers/transformers dependencies that are not in the dev extra.

[jfischoff]

@codex review