server: save processed mtmd chunks as placeholder - #27278

I reviewed the diff (mtmd get_placeholder helper + server_tokens::push_back_placeholder + the single call site in server-context.cpp). I traced how slot.prompt.tokens media chunks are consumed afterwards (cache-hit comparison via find_diff, validate, find_message_spans, slot save/restore via serialize/deserialize, slot clone()) and confirmed placeholders preserve the only fields those paths read (type, id, n_tokens, n_pos), since serialize already strips only the batch_f32 buf and keeps all metadata. Context shift is already aborted for multimodal, so the stripped chunks are never re-encoded. The change is semantically sound.

Will slow the review

(point 1) New public API in mtmd.h (mtmd_input_chunk_get_placeholder) with a single internal consumer. The only caller is server_tokens::push_back_placeholder in tools/server/server-common.cpp:514. Per the multimodal guidance, new public surface in mtmd.h is expected to be discussed first, and here it isn't obviously needed as public ABI: it exists purely to let server-common strip chunk data. Consider implementing this as an internal helper (e.g. a method on the internal mtmd_input_chunk struct, or a non-exported function in mtmd.cpp that server-common reaches through a thin existing accessor) rather than widening the public C API for one in-tree user. If it must stay public, add a one-line justification in the PR description.

Nits

(point 2) tools/server/server-common.cpp:516 aborts the whole server on placeholder failure:

GGML_ASSERT(new_chunk != nullptr && "failed to create placeholder chunk");

The previous push_back path went through mtmd_input_chunk_copy, which can only fail by throwing bad_alloc (not by aborting). mtmd_input_chunk_get_placeholder can return nullptr if the serialize/roundtrip throws (logged via LOG_ERR). A roundtrip of an internally-valid chunk shouldn't fail, but if it ever does, GGML_ASSERT takes down the server process where the old code would not. Prefer returning the error to the caller of push_back_placeholder (and ultimately send_error/slot.release) instead of asserting.

(point 3) tools/mtmd/mtmd.cpp:2338 implements "strip data" as a serialize -> deserialize roundtrip through a temporary std::vector<char>, allocated and re-parsed for every media chunk during prompt processing. The author's own comment calls it "hacky". A direct metadata-only clone (copy nx/ny/pos/image_idx/n_temporal_merge/id and batch_f32 entries with cleared buf) would avoid the extra allocation + re-parse and read more clearly. Not blocking since it is correct and the serialized form is small.

(point 4) tools/mtmd/mtmd.cpp:2316 — mtmd_input_chunk_save_impl's LOG_ERR uses __func__, which now prints mtmd_input_chunk_save_impl instead of the public mtmd_input_chunk_save. Minor loss of debuggability; consider forwarding the caller name or keeping the log at the public boundary.

(point 5) Comment wording: tools/mtmd/mtmd.cpp:2339 "faster than copy the whole batch data" -> "faster than copying"; tools/server/server-common.h:197 and tools/mtmd/mtmd.h description "doesn't contain underlay data buffer" -> "underlying".