feat: add bounded colocate transfer and Qwen4Exp converters

[dingzhiqiang]

Description

Keep transfer planning as tensor views and allocate contiguous send/receive buffers only for the active batch, including strided column slices. Copy receives back before releasing that batch. Pack compatible expert operations within configured bounds, check packing limits across inference ranks before P2P starts, and provide CUDA IPC staging allocation that restores the allocator settings used for training. The packing byte limit is not a hard cap for a single oversized tensor.

Add a reader factory so integrations can explicitly choose the bounded transport; the AWEX default remains unchanged. All participating readers must choose the same transport. This extracts the shared transport implementation from areal-project/AReaL#1731.

Own Qwen4Exp weight converters, GDN/gated-QKV layouts and frozen-parameter declarations in AWEX. Registration is explicit and accepts caller-supplied binders; checkpoint evidence and live engine ownership stay in the integration. AReaL consumes these APIs in areal-project/AReaL#1751 without a second converter implementation.

Validation

• 41 targeted CPU tests passed, including real two-rank Gloo agreement/disagreement, strided buffer allocation/release and independent GDN/QKV coordinate checks across TP sizes.
• AReaL integration regression suite: 557 passed, 9 optional-runtime skips; separate Transformers 5.16.1 vision run: 31 passed.
• Ruff check and format passed for changed files.
• GPU NCCL transfers, repeated weight equality, and cross-image CUDA IPC remain unverified and require validation before merge.

[chaokunyang]

LGTM