Add GLM5.2 DP1/EP8 load-weight slice

[xiaguan]

Summary

• Add a GLM5.2 load-weight-only crate and binary for the DP1/EP8 shape.
• Load rank0 non-expert tensors plus experts 0..31, and ranks1..7 32 routed experts each.
• Keep generation fail-closed until forward/decode lands.
• Wire the optional glm52 server feature and model detection without pulling in PP/decode/MTP runtime or new kernels.
• Optimize load time with rank-local GPU slabs, coalesced H2D copies, and CUDA-event mmap lifetime guards including error-path cleanup.

Performance

Measured on jz-38, 8x H200, real /data/models/GLM-5.2-FP8 checkpoint:

• Baseline per-tensor CudaSlice load: 81622ms
• Current implementation first run: 63420ms
• Current implementation immediate repeat: 50803ms

Validation

• cargo fmt --all --check
• git diff --check
• cargo check -p openinfer-glm52
• cargo check -p openinfer-server --features glm52
• cargo check -p openinfer-server --no-default-features --features glm52
• cargo test -p openinfer-glm52 --lib
• cargo test -p openinfer-server --no-default-features --features glm52 glm52_
• cargo build --release -p openinfer-glm52 --bin glm52_load_weights on jz-38
• ./target/release/glm52_load_weights --model-path /data/models/GLM-5.2-FP8 --tp-size 1 --dp-size 1 on jz-38, twice

Notes

• This PR intentionally excludes forward/decode/PP/MTP runtime and kernel additions.
• The load path stores raw resident weight bytes plus tensor offsets; runtime layout packing belongs to a later forward PR.
• Error-path mmap lifetime is guarded by RAII plus stream sync; happy path uses per-shard CUDA events and a bounded live mapping ring.