USB gadget: speed auto-negotiation, robust host reset/suspend handling (fixes ITX Llama boot detection)

[danifunker]

Summary

This PR makes the DWC2 USB gadget stack resilient to real-world host behavior
— full-speed-only hosts/hubs, warm reboots, and BIOS-time enumeration. The
headline features are USB speed auto-negotiation for the MSD gadget and rework
of suspend/reset handling so the device survives a host warm reboot, which
enables boot detection on the ITX Llama (and other PC BIOSes that enumerate
during POST). Addresses issue #591.

USB Auto-Negotiation

The MSD gadget previously presented a fixed high-speed configuration (512-byte
bulk packets) regardless of what the host actually negotiated, which breaks
against full-speed-only hosts/hubs.

• New CDWUSBGadget::GetNegotiatedUSBSpeed() reads the enumerated speed from
  the DWC2 device status register (DSTS.ENUMSPD, new DWHCI_DEV_STS definitions
  in dwhci.h).
• New virtual OnNegotiatedSpeed() hook, invoked from HandleEnumerationDone(),
  lets gadgets adapt their EP parameters when the negotiated speed differs from
  the configured one.
• CUSBMSDGadget overrides it: the configuration descriptor is now a
  per-instance copy (m_ConfigurationDescriptor) whose bulk wMaxPacketSize is
  rewritten to 64 (FS) / 512 (HS) on enumeration, and the EP hardware is updated
  live via the new CDWUSBGadgetEndpoint::SetMaxPacketSize() (rewrites MPS in
  the IN/OUT EP control register).
• CBW reception now requests wMaxPacketSize bytes instead of exactly 31, and
  accepts nLength >= SIZE_CBW, because hosts may pad the CBW to the packet
  boundary.

Robust suspend/reset handling (ITX Llama boot detection)

A PC BIOS enumerating during POST/warm reboot exposed several windows where
the gadget went deaf or crashed:

• Suspend is no longer treated as a disconnect. HandleUSBSuspend() previously
  fired a PnP event leading to full EP teardown and core re-init, leaving the
  device unresponsive for >100 ms — a host that resets and enumerates inside
  that window gets hardware ACKs but the software never sees ENUMDONE/SETUP. Now
  we only abort in-flight transfers, enter StateSuspended, and let the next USB
  reset re-initialize the EPs at IRQ level.
• Resets from StateSuspended/StateResetDone cancel EP0's stale SETUP transfer.
  Otherwise the next SETUP completes against the old transfer and is dropped
  with a wrong length (host reset storms, warm reboots).
• EP0 is armed for SETUP already in HandleUSBReset() (matching Linux dwc2
  behavior), closing the window between ENUMDONE and OnActivate() where a fast
  host's first SETUP was ACKed by the core but dropped with length 0.
  OnActivate() is now idempotent via the new IsTransferActive() guard.
• Malformed SETUPs no longer assert-crash. A SETUP completing with an
  unexpected length is logged, dropped, and EP0 is re-armed cleanly.
• IN EP interrupts arriving while suspended are acknowledged without
  processing instead of tripping a state assert.

Diagnostics

Added always-on LOGNOTE traces for suspend, bus reset (with state),
enumeration done (with negotiated speed), and every EP0 SETUP request — to
diagnose hosts that bounce between speeds or reset repeatedly (issue #591).

Testing

• Verified ITX Llama BIOS now detects the MSD gadget on cold and warm boot.