fix: use max_completion_tokens for gpt-5/o-series models

[mogliabr]

Problem

The Chat Completions API rejects max_tokens on reasoning models — gpt-5 and the o-series — returning 400 Unsupported parameter: 'max_tokens'. These models require max_completion_tokens instead.

buildPayload() currently always sends max_tokens when maxTokens is set:

$payload['max_tokens'] = $options['maxTokens'];

Fix

Select the token-limit field based on the model family prefix:

$payload[$this->tokenLimitKey($model)] = $options['maxTokens'];

tokenLimitKey() returns max_completion_tokens for gpt-5, o1, o3, o4 (matched by prefix, so future gpt-5.x releases are covered without changes) and max_tokens for everything else.

Tests

Added coverage for gpt-5.6-luna and o3-mini asserting max_completion_tokens is emitted and max_tokens is absent. The existing gpt-4-turbo test still asserts max_tokens, confirming classic chat models are unaffected.

composer lint and composer test (71 passed) pass locally.

[mogliabr]

I get this bug using lib my personal fix can be affected by furure updates so i sended the path. Thanks for created it. Fist time using github im only subversion and cvs user for commit. Get learn how use git hub.