fix: DeepSeek-OCR server crash, template routing, and CUDA OOM fallback

[soymh]

Upstream PR: #17400 Original implementation by sfallah (sf/deepseek-ocr branch):
https://github.com/sfallah/llama.cpp

Co-authored-by: sfallah sfallah@users.noreply.github.com

Fixes:

• server crash: GGML_ASSERT(batch.n_tokens > 0) when mtmd image processing consumes all prompt tokens (inject synthetic token before assertion)
• server crash: slot.task NULL dereference after release() on mtmd OOM
• server crash: ggml_backend_sched_alloc_graph segfault when CUDA OOM (check return value, matching sfallah's upstream guard pattern)
• template routing: --chat-template deepseek-ocr was rendered as Jinja text instead of resolving to the built-in template (auto-detect + legacy fallback)
• bitmap/marker mismatch: get_media_marker() returned random string instead of mtmd_default_marker(), causing tokenizer to split on wrong marker
• CUDA OOM: tensor loading falls back to CPU backend when GPU allocation fails
• mmproj GPU: -ngl 0 now also disables mmproj GPU (matches user intent)

Overview

Additional information

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure:

[ngxson]

there changes are too invasive, break all other models, we cannot accept

[ngxson]

ref: #23345