Fix potential data loss when writing to congested L2CAP CoC.

Observed this rare problem with busy sockets on Wear OS watches paried
with iOS:
1. Multiple poll events are received on the same poll call.
2. While processing commands ts[h].poll_count is incremented without updating pfds.
3. process_data_sock reads stale pfds entries and processes POLLIN on a
  fd that didn't actually receive this event.
4. Data is read from fd but dropped because the channel is congested.

The change copies the initial poll_count to avoid reading stale pfds entries.
Verified Wear iOS comms stack does not miss packets when using L2CAP
CoC after the change.

Bug: 223974328
Tag: #stability
Test: gd/cert/run

Change-Id: I676aa9f2f1074e5df89bbbf08f986c5fe8972358