perf(wan2.2): Optimize attention kernels, rematerialization, and Ulysses parallelism for TPU v7x

- Implement custom Pallas splash & ring attention forward and fused backward kernels
- Optimize rematerialization policy with custom checkpointing rules
- Implement Ulysses sequence parallelism with 2-phase all-to-all communications
- Eliminate device-to-host sync stall via deterministic host-side expert routing
- Support multi-host evaluation metric gathering with multihost_utils.process_allgather
- Specify sequential dimension semantics in Pallas backward kernels for safe accumulation

Author: Toshi Pahadia