if_bnxt: Fix HWRM mailbox/DMA teardown race on detach

[sumit.saxena_broadcom.com]

Summary

iflib's generic device-deregister path never drains the admin task
before calling IFDI_DETACH, so bnxt_update_admin_status() (scheduled
once/sec) could still run concurrently with bnxt_detach(), racing on
softc->hwrm_lock and the shared HWRM request/response DMA buffer that
bnxt_detach() destroys. That race could leave the firmware-side HWRM
mailbox inconsistent, surfacing as "Timeout sending HWRM_VER_GET" on
the next module load.

Add a softc->detached guard: set it as the first statement in
bnxt_detach(), and bail out of bnxt_update_admin_status() immediately
when it's set.

[sumit.saxena_broadcom.com]

AI scan fixes:

• softc->detached was a plain bool, written from bnxt_detach() and polled from bnxt_update_admin_status(), which run on different threads with no lock in common at those two points. A plain bool gives no cross-thread visibility or ordering guarantee, so the reader could still observe a stale false well after the writer's store, narrowing the race this commit describes rather than closing it. Change the field to volatile int and access it via atomic_store_rel_int()/atomic_load_acq_int() at both sites.

[gallatin]

Note that this may no longer be needed as iflib_device_deregister() drains the private taskqueue before IFDI_DETACH since commit ba353c8950d5