fix(iroh): preserve relay receive outputs and retry after message drops

[Theodus]

Description

Fixes #4572.

RelayTransport::poll_recv indexed output slots by receive attempts. Dropping an oversized batch could leave a gap, causing the caller to miss a later usable message. If every attempt dropped data, the poll could also return Pending without requesting a wake for traffic still in the queue.

This change:

• Index bufs, metas, and recv_infos by num_msgs, keeping usable outputs contiguous after drops.
• Use continue instead of break after dropping an oversized single datagram.
• Call cx.waker().wake_by_ref() when message drops exhaust the bufs.len() attempt limit, ensuring another poll.

API Changes

None.

Notes & open questions

The regression tests use 1,472-byte receive buffers and cover both oversized batches and oversized single datagrams:

• One verifies that usable messages occupy consecutive output slots with the correct payloads, strides, and source addresses.
• The other verifies the receive-attempt limit, a retry wake when message drops exhaust that limit, and no self-wake on an empty queue.

These tests replace the narrower oversized-batch progress test.

Change checklist

• Self-review.
• Documentation updates following the style guide, if relevant.
• Tests if relevant.
• All API changes documented.
• This PR was created by a human that thought critically about the
  proposed change and wrote as clear and concise a description as
  they could.
• This PR isn't slop, and is carefully crafted to have the
  intended effect.