bledev: fast reconnects, and REPL output that arrives intact

[bdbarnett]

Two v1 cleanups for #81, stacked on bledev-host. Merge #78 and #81 first.

The reconnect flake. The REPL gate once failed one run in five: bleak couldn't subscribe on a connection made 1.5 s after the last one closed. tests/bledev_board/reconnect_loop.py (log in, run a line, close, wait 1.5 s, from the laptop) caught it at about one cycle in fifty, always the same shape: bleak's connect returned in 70-440 ms instead of the usual 1.7 s or more, Windows said the device was connected, the board never saw a connection (its IRQ log shows it advertising the whole time), and the first GATT operation hung until Windows gave up nine seconds later.

• Against bledev-host's repl.py on the LCD-7: 7 failures in 465 cycles (5/315 earlier today, 2/150 as a re-check just now).
• Against this branch's repl.py: 0 in 270 (120 with the retry available and never needed, 150 with it switched off).

So the board side was the cause, and the fix is the REPL change below. I haven't found how the old timer handling made Windows believe in a connection the board never saw, so bledev.connect_and_set_up() stays as a defence: it bounds setup at 4 s and retries up to three times. nus.connect() uses it, and midi.connect() does on #85. The contract checks both the retry and giving up, and fails with the retry planted out.

REPL output. Printing a 16,890-character list over the BLE REPL came back corrupt 5 times out of 5: the retry timer's callback could land inside a flush() that print() had started, so a chunk went out twice and the next never did. A re-entry guard fixed that. Then the board panicked twice in about 45 long prints (InstrFetchProhibited in esp_timer's task): re-arming a one-shot machine.Timer(-1) while its alarm is being dispatched makes the esp32 port call a NULL handler. The REPL now starts one periodic timer and never re-arms it. After both fixes: 60 long prints intact, no panic, with the console captured throughout. The REPL gate now includes the long print.

Fake contract 43/43 on CPython and unix MicroPython; test_bledev passes, including every plant.

Not done: the esp32 machine.Timer re-init race is upstream MicroPython's (ports/esp32/machine_timer.c: machine_timer_deinit() sets handler = NULL while esp_timer may be dispatching it). bledev no longer triggers it. No upstream draft was written for it in this PR.