server: restore schema-declared argument types for GLM tool calls

[solodddd]

Fixes #569.

Problem

On the glm5.2 branch, the GLM tool-call parser stores every <arg_key>/<arg_value> argument as a JSON string ("max_results":"10", "exclude_domains":"[]"), because GLM's wire format carries no type information — unlike DeepSeek's DSML string="true|false" attribute. Clients that validate tool-call arguments against the tool's JSON schema (Chatwise and anything zod/ajv-based) reject every call with a non-string parameter.

Fix

Schema-aware type restoration, no heuristics:

• While parsing the request's tool schemas, capture each property's declared "type" alongside the existing property-order bookkeeping (tool_schema_order gains a parallel prop_type array).
• At client-serialization time, a string argument whose declared type is number/integer/boolean/array/object/null and whose text is exactly one valid JSON value of that kind is emitted raw; everything else (declared strings, undeclared properties, values that don't match their declared type, untyped/union schemas) keeps the current behavior.
• Applied at all client-facing outputs: chat completions (final + streaming deltas), Responses API items/deltas, and Anthropic tool_use blocks (final + SSE flush). Prompt re-rendering paths are untouched. DeepSeek DSML calls are unaffected (their values are already typed; the coercion only ever unquotes a string whose schema says non-string).

Tests

Added test_openai_tool_args_restore_schema_types covering: number/boolean/array restoration, declared strings staying strings, undeclared properties staying strings, and a value that is not valid JSON of its declared type staying a string.

Per CONTRIBUTING:

• Machine/backend: Apple M3 Ultra (512 GB), Metal backend, macOS.
• Model quant: GLM-5.2 routed Q4_K (GLM-5.2-UD-Q4_K_RoutedQ4K.gguf).
• make — clean build, no new warnings.
• ./ds4_test — all tracks pass (including --server, which covers the tool-call parsing/serialization touched here).
• Live verification against a running server, non-streaming and streaming:

Before:

"arguments": "{\"query\":\"...\",\"max_results\":\"10\"}"

After (same request, tool schema declares max_results: number, exclude_domains: array):

"arguments": "{\"query\":\"...\",\"max_results\":10,\"exclude_domains\":[\"example.com\"]}"

No inference backend is touched — the change is confined to request-schema parsing and response serialization — so there is no speed-regression surface.

Developed with the assistance of Claude (Anthropic), reviewed and tested by me.