Don't buffer files in memory at send time

[PaulSonOfLars]

What

Addresses #259

Previously, we would pipe files while creating the multipart reader.
This changed in #242, where i opted to avoid double-marshalling fields - but introduced a regression where files were double-buffered in memory in an attempt to maintain the fix from #231.

This PR introduces regression tests to cover those edge cases, and then updates the requests code to correctly use pipes and allow retries without double-buffering files.

Impact

• Are your changes backwards compatible? Y
• Have you included documentation, or updated existing documentation? Y
• Do errors and log messages provide enough context? Y

[danielwaichcesl]

Hi @PaulSonOfLars,

Tested PR #260 with a 31 MB file upload - works perfectly! 🎉

Results
Memory usage:

Before: ~66 MB peak (buffered entire file)
After PR #260: ~8.8 MB peak
87% memory reduction ✅
Test details:

File: 31 MB
Upload: successful in ~17s
No memory leaks
GC reclaims all memory correctly
Impact
This solves our issue with large file uploads in containerized environments. The streaming implementation works great - files are no longer buffered in memory.

Thank you for the quick fix! 👏