fix: convert adaptive thinking to enabled for Bedrock models that don't support it

[blinkagent]

When Claude Code sends thinking.type: "adaptive" (the default for Opus 4.6 / Sonnet 4.6), Bedrock models that predate adaptive thinking support (e.g. Sonnet 4.5 in GovCloud us-gov-west-1) reject the request with a 400 error:

thinking: Input tag 'adaptive' found using 'type' does not match any of the expected tags: 'disabled', 'enabled' 

Changes

intercept/messages/base.go

• Added bedrockModelSupportsAdaptiveThinking() to detect whether the target Bedrock model supports adaptive thinking (only Opus 4.6 and Sonnet 4.6 do)
• Added convertAdaptiveThinkingForBedrock() which converts {"type": "adaptive"} to {"type": "enabled", "budget_tokens": 10000} when the target model doesn't support it
• Extended augmentRequestForBedrock() to call the new conversion after the model name remap

intercept/messages/base_test.go

• Added TestBedrockModelSupportsAdaptiveThinking — 11 cases covering all model name formats
• Added TestConvertAdaptiveThinkingForBedrock — 6 cases: conversion for non-4.6 models, preservation for 4.6 models, no-op for enabled/disabled/absent thinking

Fixes #219

[uzair-coder07]

I have used blink to create this PR. I will take a look at the code and will test it out before I ask for a review.

[uzair-coder07]

I had a first pass through the code and I believe hardcoding the budget_tokens to 10000 can backfire since Anthropic API requires budget_tokens < max_tokens? Maybe we should be checking the request's max_tokens. If a client sends a request with max_tokens: 8000 (or less), Then we might run into another 400 error like "max_tokens must be greater than thinking.budget_tokens".