glm5-next: fall back to the deepseek4 spelling for the k-pool tensors

[danielhanchen]

Follow up to my comment on ggml-org#27773. I originally wrote this as a loader change and have cut it back to a single model file, since none of it needs to touch core.

GGUFs for this model were published before the naming settled, so two conventions are in the wild:

	general.architecture	k-pool tensors
this PR	glm5-next	blk.N.indexer.kpool_{gate,ape}
ggml-org#27754, ggml-org#27752	glm5next	blk.N.indexer_compressor_{gate,ape}

What this changes

Only the tensor half, and only in src/models/glm5-next.cpp: try the canonical name first, fall back to the deepseek4 spelling. It reuses the existing LLM_TENSOR_INDEXER_COMPRESSOR_{WGATE,APE} entries, so there is no new enum, no new name table entry, and nothing outside this file changes. LLM_TN_IMPL::str() is a plain enum to name lookup with no arch check, so this is just asking for a name that is already in the table.

The architecture string and the 40 arch-prefixed KV keys are deliberately not aliased. Those all live in shard 1, which is 9.4 MB with zero tensor infos, so the publishers on the other spelling can rewrite that one small file instead of asking you to carry a compatibility path for it. That is our side's job, not yours.

Verification

Published unsloth/GLM-5.3-Flash-GGUF UD-IQ1_S, shard 1 rewritten to glm5-next in 0.338 s, both data shards left byte for byte as uploaded and still containing blk.N.indexer_compressor_*:

print_info: arch                  = glm5-next
load_tensors: offloaded 47/47 layers to GPU
load_tensors:        CUDA0 model buffer size = 85766.72 MiB
[ Prompt: 53.2 t/s | Generation: 61.7 t/s ]

Without the fallback the same file fails with check_tensor_dims: tensor 'blk.3.indexer.kpool_gate.weight' not found.

No other architecture can be affected, since the change is inside the GLM5-Next branch of one model file.

One note on the fallback path: LLM_TENSOR_INDEXER_COMPRESSOR_APE is registered GGML_OP_GET_ROWS while LLM_TENSOR_INDEXER_KPOOL_APE is registered GGML_OP_ADD. Yours is the correct one, since the tensor feeds ggml_add. That field only feeds select_weight_buft, so the effect is buffer placement on one small tensor and never numerics, but it is worth knowing that files taking the fallback get the older registration.