Document the Buffer/Batch constraints on tower_fallback

[teor2345]

Is your feature request related to a problem? Please describe.

tower_fallback::service and tower_fallback::future implement some complex poll_ready/call state handling between two services, which is particularly risky if either of those services are Buffered or Batched. This state handling happens across both modules.

The current code appears to be correct, but the Buffer/Batch constraints are undocumented. It also depends on the modules remaining in sync.

Here are the underlying issues:

• If lots of tasks with data dependencies use a Buffer or Batch with a small bound, Zebra can hang.
• Buffers or Batches with small bounds can also limit concurrency, particularly if the underlying service does a lot of work in its call.

Describe the solution you'd like

• Document the internal states of tower_fallback::future
• Document how tower_fallback avoids these risks with Buffered or Batched services

Describe alternatives you've considered

Do nothing. If callers use tower-fallback incorrectly, they may hang.

Additional context

See #1593 for the overall Buffer and Batch hang issue, analysis, and design patterns.

[teor2345]

This task has been revised based on the fixes to the underlying Buffer issue, see #1593 for details.

[teor2345]

The current settings seem fine.