feat(bedrock): support token counting

[aidand-ant]

Amazon Bedrock now supports token counting via the CountTokens API, so the SDK no longer needs to reject count_tokens() calls on the Bedrock client.

What changed

• _prepare_options now rewrites POST /v1/messages/count_tokens (and the ?beta=true variant) to Bedrock's POST /model/{modelId}/count-tokens endpoint. The Anthropic-format request body is wrapped in Bedrock's CountTokensInput shape as a base64-encoded invokeModel body. The model ID is URL-quoted the same way as for create, so inference profile ARNs work.
• BaseBedrockClient gains a _process_response_data override that maps Bedrock's {"inputTokens": N} response to {"input_tokens": N}, so callers get the same MessageTokensCount type as the first-party API.
• The Bedrock beta messages resource previously only exposed create; it now also exposes count_tokens (including raw/streaming response wrappers).

Usage

from anthropic import AnthropicBedrock

client = AnthropicBedrock()
count = client.messages.count_tokens(
    model="anthropic.claude-3-5-sonnet-20241022-v2:0",
    messages=[{"role": "user", "content": "Hello, world"}],
)
print(count.input_tokens)

Tests

Added sync, async, and beta tests covering the URL rewrite, the base64-wrapped request body, and response parsing. All AWS-related tests pass (139 passed).