feat(server): add Poolside V1 reasoning and tool parsers

[pedro-moser]

Summary

The Laguna architecture added in #102 uses Poolside's implicit-open reasoning format and GLM-style tool-call tags, but automatic parser selection currently falls through to unrelated generic parsers. As a result, Laguna can load and generate while reasoning/tool transport is silently misclassified.

Expected behavior

For Laguna / Poolside checkpoints, --reasoning-parser auto and --tool-call-parser auto should select a dedicated poolside_v1 pair:

• reasoning starts implicitly and closes at </think>
• tool calls use <tool_call>, <arg_key>, and <arg_value> tags
• streaming must tolerate arbitrary tag chunk boundaries
• JSON-schema-declared strings preserve source-code whitespace; typed non-string arguments retain GLM conversion behavior
• explicitly selecting poolside_v1 must be accepted by both CLI choices

Proposed implementation

• Add poolside_v1 to the reasoning and function-call parser factories.
• Reuse the existing GLM 4.7 state machine for tool tags while preserving leading/trailing whitespace only when the property schema is exactly type: "string".
• Auto-select the parser for Laguna architecture/path markers.
• Keep existing glm47 buffered and streaming behavior unchanged.

Validation

On a branch containing current main plus #102:

• 254 parser/streaming tests pass across the auto-selection matrix, Poolside-specific cases, GLM/Qwen/Minimax/Muse parsers, default/disabled thinking modes, and reasoning models.
• Poolside-specific coverage includes buffered and arbitrary-chunk streaming, escapes, text after tool calls, enum-only/untyped schemas, exact string whitespace, and schema-aware GLM coercion for explicitly typed non-string arguments.
• A real unsloth/Laguna-S-2.1-GGUF revision 750f92f90cf54159c4d7a610cb7b3e74498e75c6 run auto-selected both parsers without explicit flags.
• Real buffered and streaming requests produced correctly transported tool calls when the checkpoint emitted a valid closed tool block. Failures from the IQ1 checkpoint emitting a tool block inside unclosed reasoning are model-generation failures, not parser corruption.

Related: #102