fix(files): infer document upload content types

[jeet-dekivadia]

Summary

• infer application/pdf and text/plain for multipart files with .pdf and .txt filenames
• keep explicit Anthropic::FilePart#content_type values authoritative
• preserve the existing application/octet-stream fallback for unknown formats and add focused regression coverage

Addresses #128.

Validation

• git diff --check
• verified File.extname handling for Pathname, missing filenames, and uppercase extensions
• added util regressions for PDF inference, text inference, and explicit MIME override precedence

Remote Linux validation

• git diff --check
• bundle install
• ./scripts/lint
• ./scripts/test

[jeet-dekivadia]

Small follow-up while this is waiting for review. The content-type inference only fills in common document uploads by filename extension when the caller has not already supplied an explicit Anthropic::FilePart#content_type. Explicit content types stay authoritative, and unknown extensions continue to use the existing octet-stream fallback.