Pin unslothai#116 (GLM-5-Next)

[danielhanchen]

Adds #116 (GLM-5-Next / GLM-5.3-Flash) to the nightly pin set, via the carry in #118.

The pin points at the carry, not at glm5next/public. The PR branches off fork master, so its head carries 77 fork CI paths including pr-set.json itself; #118 is b10632 plus the PR's 42 paths and nothing else, byte-identical to the PR head on all of them.

#118 is built on top of the qwen4exp carry rather than on b10632 directly, because those two pins are not independent: both add a sparse-attention indexer cache and both declare the same filter_idx in create_memory, so a side-by-side merge declares it twice and the build stops. That composition is resolved once in #118 rather than attempted nightly. The consequence is recorded there: dropping the qwen4exp pin no longer drops its code.

Placed last, after the qwen4exp pin it depends on.

Verified against b10632:

• all seven pins merge in list order; six clean, #70 via additive_merge.py, no hand resolution in the nightly path
• the merged tree compiles clean, 573/573 targets, with the preflight gate's config plus LLAMA_BUILD_TESTS=ON
• test-llama-archs exits 0; glm5next OK, qwen4exp OK, inkling's documented skip intact
• 3766b41229 is in Carry unslothai#116 (GLM-5-Next) composed onto the qwen4exp b10632 carry #118's commit list, so the pin membership gate passes
• mirrored to refs/pins/3766b41229c20249fd4d83d7ba297499d50e9b80

Merge #118 first, or at least do not delete its branch while this pin is listed.