common : add support for multiple end sequences in the reasoning budget sampler

[aldehir]

Overview

Add support for multiple end sequences in the reasoning budget sampler. This is required to support #24202, which allows Qwen 3+ models to terminate their thinking process with either </think> or <tool_call>.

This PR contains quite a bit of changes to support this:

• The trie/ac data structures in peg-parser.cpp are moved out to a separate trie.cpp/h. These are used to handle the token matching underneath, as they already support matching multiple sequences.
• Refactored reasoning_end_tag to reasoning_end_tags.
• Updated all parser implementations to return a list of tags, although currently they all only provide one.
• When forcefully terminating reasoning, the first sequence is used. We could retain a single reasoning_end_tag and have an extra reasoning_additional_end_tags field if that feels more appropriate.
• Added common_reasoning_budget_get_end_match() to obtain the matching sequence. Since <tool_call> is also a trigger token for the grammar sampler, we need to feed it to the sampler when grammar_lazy = true. Either this, or we revert our decision to suppress the grammar sampler while reasoning.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, to extract the data structures and write some tests.