Bump tokio-rustls to 0.26.6

[rdettai-sk]

Description

Bumps tokio-rustls from 0.26.4 to 0.26.6 (lockfile only). The change that matters is tokio-rustls#198, released in 0.26.5: poll_read keeps decrypting records until hyper's buffer is full, instead of returning a single ≤16 KB record per poll. 0.26.6 fixes a 0.26.5 regression in which ConnectionAborted became a clean EOF.

How was this PR tested?

Test environment:

• S3 connection task polls dropped from ~70k to ~7k/s per pod, and bytes per poll went from ~16 KB to ~190 KB.
• Tokio time per downloaded MB did not change measurably (~2.2 ms/MB before and after). The per-byte cost is kernel receive work, not per-record userspace overhead.
• Because polls are fewer and longer, the 16-256 ms tail of S3 task polls grew (0.15 → 2.5/s per pod), during a period when CFS throttling in the test environment also rose. Upstream draft tokio-rustls#209 caps the bytes read per poll to address this.

🤖 Generated with Claude Code

[Copilot]

Copilot wasn't able to review any files in this pull request. Check if the Files changed in this pull request are included in default exclusions.

[rdettai-sk]

The result of this bump is not conclusive. We should work on this upstream first (rustls/tokio-rustls#209)