fix(arm): do not clear STOPF on the I2C NACK path

[nehalkpatel]

WaitForFlag cleared NACKCF and STOPCF together when it saw a NACK. But
autoend generates the STOP itself in response to that NACK, and
FinishTransfer is what waits for the resulting STOPF — so clearing it there
left that wait with nothing to observe, burning the full 25 ms transfer timeout
on every failed transfer.

Not a hang: the timeout bounds it, and the error the caller sees is the NACK's
kOperationFailed rather than the discarded kTimeout. But it made a missing
device cost 25 ms per attempt instead of the ~90 µs an address phase takes at
100 kHz.

Verified on hardware, not just reasoned about

Flashed i2c_demo to the F767ZI with nothing on the bus and broke on both
failure branches of WaitForFlag to see which one the driver actually takes:

Breakpoint 1, mcu::(anonymous namespace)::WaitForFlag (registers=0x40005400, flag=2)
    at src/libs/mcu/arm_cm7/i2c.cpp:82
#0  WaitForFlag ... i2c.cpp:82          <- the NACK branch
#1  mcu::I2CBus::SendData (address=80)  ... i2c.cpp:167
#2  app::I2CDemo::Run                   ... i2c_demo.cpp:41
ISR now = 0x00000031  (NACKF=1)

ISR = 0x31 is the whole argument for this change: TXE (bit 0), NACKF
(bit 4) and STOPF (bit 5) are set at the same instant. The STOP has already
happened by the time the NACK is detected — so the old code was clearing a flag
FinishTransfer was about to wait for, guaranteeing the timeout every time.

Incidentally this also confirms the I2C driver end to end on real silicon: the
address goes out at 0x50, nothing answers, the peripheral raises NACKF, and
SendData returns kOperationFailed without hanging.

Checks

• Host: 32/32
• ARM: Debug and Release build clean, 6 images pass tools/verify-firmware.sh
• tools/format.sh --check clean

🤖 Generated with Claude Code