fix(layers/retry): do not reopen a read once the range is exhausted

[swanandx]

When a stream breaks after its last byte, the layer advances the range to size zero and reopens with it. Nothing is left to fetch, and an empty range has no valid Range header.

Return an empty buffer instead, which callers treat as end of stream.

An unbounded read has no size to compare against, so it still reopens past the end. Catching that needs the object size.

Which issue does this PR close?

Closes #8342

Rationale for this change

RetryReadStream::read advances the range by what it has read before reopening. If the stream breaks after the last byte, the range is already empty by then, and it reopens with it. Nothing is left to fetch.

With a LoggingLayer in the stack that panics, because the layer calls to_string() on the range and Display returns Err for a zero size. ErrorContextLayer does the same on its error path. Without either, the request goes out with an empty Range header.

What changes are included in this PR?

Return an empty buffer instead of reopening. Callers already treat that as end of stream; the trait's own read_all does (core/core/src/raw/oio/read/api.rs:119).

This covers bounded ranges and suffixes. An unbounded read has no size to compare against, so it still reopens past the end. Catching that needs the object size.

Adds a test that fails without the guard. test_retry_read reads unbounded, so it still reopens and still expects five attempts. It is unchanged.

Are there any user-facing changes?

Don't think so

Breaking changes

No

AI Usage Statement

• Harness: Claude Code
• Model: Claude Opus 5
• Effort: default; the harness does not surface a setting
• Role: helped trace the reopen path and draft the test.