Support DeepSeek-R1 Qwen

[cebtenzzre]

This PR adds DeepSeek-R1 Qwen support by:

• Rebasing llama.cpp on a slightly newer upstream commit (ggml-org/llama.cpp@a39ab216a from Oct 2 instead of ggml-org/llama.cpp@95bc82fbc from Sep 26)
• Cherry-picking the one-line change required to load the DeepSeek GGUF
• Adding a prompt template substitution that doesn't rely on things like namespace and str.split which Jinja2Cpp doesn't implement
• Implementing a regex_replace function compatible with others seen in the wild to support the necessary text manipulation in the DeepSeek template.

Tested on DeepSeek-R1 Qwen 7B. Will update with results with 14B and 32B soon.

Note that the replacement prompt template removes the {{ bos_token }} from the Jinja template to avoid a double BOS, since at least the bartowski GGUF I tried specifies tokenizer.ggml.add_bos_token of true already.

[cebtenzzre]

The 14B and 32B models also work as expected. All three models (7B, 14B, 32B) work on both CUDA and Kompute.
edit: Also tested 1.5B successfully.

[WuShichao]

How about Vulkan backend? I can ran CUDA with DeepSeek-R1-Distill-Llama-8B-Q4.