py/gc: Support multiple heaps. - #3533

The PicoMite tree ran its TinyUSB 0.21 host work on this same hardware -
two keyboards (one marginal), a touch panel and a flash drive behind one
hub - across software, hardware and power-on resets, and wrote up what
actually held (docs/usb-host-hardening).  Root cause: the RP2 SIE keeps
ONE handshake-result latch shared between EPX and the interrupt-endpoint
poller (upstream micropython#3533), so an interrupt poll can overwrite a control
transfer's ACK before the IRQ handler reads it, and the transfer is
misread as RX_TIMEOUT.  The validated answer is to tolerate the race
above the silicon, not to adopt the strict micropython#3533 driver rewrite - which
enumerated FEWER devices on the marginal rig.  Applied here:

  hcd_rp2040.patch  Grace period for EP0 RX timeouts: within a 1 s
                    window the transfer is left armed and the clobbered
                    result outlasted; the window closes on any EP0
                    completion.  Expiry (a genuinely dead device) takes
                    the original fail path, which keeps our buffer clear
                    (micropython#3874).  Bulk/interrupt keep the fast-fail: their
                    flow control is NAK, which never raises RX_TIMEOUT.

  usbh.patch        Enumeration-exclusive control dispatch: while a
                    device enumerates, only address 0, the enumerating
                    address and the hub in use may claim the control
                    slot; LED writes and touch handshakes wait in the
                    pending FIFO (gated at the claim, the two drain
                    checks and the dispatcher, by peeking the head).
                    And a failed enumeration now disables its hub port
                    (CLEAR_FEATURE(PORT_ENABLE), async no-op callback) -
                    0.21 left the abandoned device enabled at address 0,
                    where it answers in parallel with the next device's
                    bring-up.  The 100 ms reset recovery (micropython#3876) stays.

  mp_usbh.c         Mount-callback prints deferred: the callbacks now
  usb_msc.c         format into a small static ring and mp_usbh_task()
                    prints it after tuh_task() returns.  A print in the
                    callback runs the VM via dupterm mid-enumeration -
                    the measured cost of "one connect chime" on the
                    PicoMite rig was the marginal device.  The existing
                    reentrancy guard treated the symptom; this removes
                    the stall itself.

  tusb_config.h     CFG_TUH_CONTROL_PENDING_QUEUE_SZ 4 -> 8: the gate
                    parks more traffic in the FIFO during enumeration.

Not adopted, by the doc's own measurement: the micropython#3533 reference hcd
(fewer devices on marginal hardware; revisit when merged upstream).
Already carried, now cross-validated: MULTI_HUB_FIX and the 64-entry
event queue.  Builds clean; the board leg - all reset kinds, hot
attach, MSC beside HID, pulling the marginal device - is owed before
this is believed.

Co-Authored-By: Claude Fable 5 <[email protected]>