fix(azure): match the generation, not one release, for max_completion_tokens

[ntdatt812]

Closes #12981.

azure/gpt-6-astra fails validation and every request:

'max_tokens' is not supported with this model. Use 'max_completion_tokens' instead.

Why

AZURE_COMPLETION_TOKEN_DEPLOYMENT matched a literal gpt-5:

/(?:^|[/_-])(?:gpt-5|o(?:1|3|4))(?:[._-]|$)|^gpt-chat-latest$/i

So the max_tokens → max_completion_tokens swap covered GPT-5 and the o-series and nothing after them. The requirement is a property of the generation, not of one release — gpt-6-astra rejects max_tokens for exactly the reason gpt-5 does — so each new family arrives broken and needs another patch. Matching the range fixes GPT-6 and whatever follows in the same edit.

Why a range and not gpt-\d+

Azure's own deployment name for GPT-3.5 is gpt-35-turbo — no dot. A digit-run matches it, and the swap would then strip the max_tokens that deployment actually requires. Breaking a working deployment is worse than the bug being fixed, so the range stops at 19: gpt-(?:[5-9]|1\d). That still leaves room for a future gpt-10.

There is a test for exactly that, and it is the test that fails if someone later widens the pattern.

Verification

tests/unit/azure-param-rules.test.ts — 10 passed. Each claim checked by reverting it:

reverted	fails
back to the gpt-5 literal	only generations after GPT-5 convert max_tokens too (#12981)
widened to gpt-\d+	only gpt-35-turbo is not a GPT-3.5 deployment caught by the generation range

The pattern lives in one place and has one consumer (applyAzureParamRules, shared by azure-openai and azure-ai), so both wire paths pick this up.