Fix slow tokenize() on Windows/MSVC

[aldehir]

I was testing ggml-org/llama.cpp#18353 and test-chat is unbearably slow on MSVC (always has been).

The cause is MSVC's std::regex engine. The regex_search() calls take a considerable amount of time during tokenization. I added some profiling and tested with the Apertus-8B-Instruct.jinja template:

[Profile] parseExpression: 4.93307 s (102 calls)
[Profile] non_text_regex: 0.0153389 s (319 calls)
[Profile] block_keyword_regex: 0.0048339 s (216 calls)
[Profile] block_close_regex: 0.0068446 s (216 calls)
[Profile] block_open_regex: 0.0817274 s (535 calls)
[Profile] expr_close_regex: 0.0035717 s (102 calls)
[Profile] expr_open_regex: 0.0604591 s (637 calls)
[Profile] comment_regex: 3.11919 s (639 calls)
[Profile] tokenize: 18.1967 s
Parsed template from: .\Apertus-8B-Instruct.jinja

18.2s to parse.

This PR adds std::regex_constants::match_continuous to the regex_search() calls. This flag instructs the regex engine to only match from the start of the iterator, which aligns with the existing match.position() == 0 check.

[Profile] parseExpression: 0.0225291 s (102 calls)
[Profile] non_text_regex: 0.013297 s (319 calls)
[Profile] block_keyword_regex: 0.0039072 s (216 calls)
[Profile] block_close_regex: 0.0052241 s (216 calls)
[Profile] block_open_regex: 0.0081572 s (535 calls)
[Profile] expr_close_regex: 0.0025207 s (102 calls)
[Profile] expr_open_regex: 0.0064239 s (637 calls)
[Profile] comment_regex: 0.0076846 s (639 calls)
[Profile] tokenize: 0.117085 s
Parsed template from: .\Apertus-8B-Instruct.jinja

0.12s, much better.

[CISC]

The model download CI failures are a PITA, but LGTM.