Fix Bedrock token count and IDs for Anthropic models

[michaelnchin]

Fixes #314

The get_num_tokens() and get_token_ids() methods of BedrockLLM and ChatBedrock currently fail when used with Anthropic models and anthropic>=0.39.0 installed.

In such scenarios, this PR fixes get_num_tokens() and get_token_ids() by falling back to the base class implementations in BaseLanguageModel and in BaseChatModel, which use the HuggingFace GPT2 Tokenizer.

If anthropic<=0.38.0 (and other requirements) are present instead, the Anthropic SDK token methods will continue to be used as normal.

Note about tokenizer accuracy:

The GPT2 and Anthropic SDK tokenizers (see here) both may produce inaccurate token estimates for Claude 3 and 3.5 models. To obtain a more accurate estimate, Anthropic recommends using the new Count Message Tokens API. For example:

import os
import anthropic

os.environ["ANTHROPIC_API_KEY"] = "<your-api-key>"
anthropic.Anthropic().messages.count_tokens(
    model="claude-3-5-sonnet-20241022",
    messages=[
        {"role": "user", "content": "Hello, world"}
    ]
)

As another alternative, you can implement your own token counter method, and pass this using custom_get_token_ids when initializing the model. This will override both the GPT2 and Anthropic tokenizers.

[3coins]

@michaelnchin
Thanks for submitting these changes. A question about the version checks, and a suggestion to use of warnings rather than logger statements.