Fix ESP32 I2C transaction locking

[ainyan03]

What changes

Lock leak on an invalid address. beginTransaction() took the port lock and then let restart() reject an out-of-range address. The state stayed state_disconnect, so the following endTransaction() returned early without unlocking or restoring the saved registers. Every later transfer on that port then blocked forever (the same task deadlocks on the non-recursive mutex). The address is now checked before the lock is taken.

Transaction owner. Rejecting before the lock means a failed begin no longer holds the port, so the cleanup after it must not touch a transfer that another task is running. The hardware port now records the task that owns the open transaction:

• endTransaction(), restart(), writeBytes() and readBytes() return mode_mismatch when the calling task does not own the transaction. This also covers a task that keeps writing after an error has already released the port.
• A second beginTransaction() from the owning task returns mode_mismatch instead of deadlocking.
• transactionWrite() / transactionRead() / transactionWriteRead() no longer call endTransaction() when beginTransaction() failed.

Contract: do not call endTransaction() after a failed beginTransaction().

Atomic bitOn() / bitOff(). writeRegister8() with a mask read and wrote the register in two transactions and released the lock in between, so a concurrent update of the same register from another task could be lost. The port mutex is now recursive and held from the read to the write. The bus sequence is unchanged (a STOP between the read and the write), so devices see the same traffic as before. This serializes callers of this library only. Another driver sharing the bus can still transfer between the two, and software (bit-bang) ports keep the previous behavior.

Checked

• Hardware, internal I2C bus of each board: beginTransaction(port, 0x800) + endTransaction(), then a register read of an on-board device from another task. "Held" runs the same while another task keeps an open transaction on the port.

Board	Chip	Port	Before: read after the failed begin/end	After: idle	After: held
Core2 v1.3	ESP32	HP 1	never returns (lock leaked)	returns, value correct	other task's transaction completes, stray endTransaction() returns mode_mismatch
AtomS3R	ESP32-S3	HP 1	—	same	same
ToughC5	ESP32-C5	LP	—	same	same
Tab5	ESP32-P4	HP 1	—	same	same

Bus scans, touch, PMIC/RTC and the IO expanders keep working with the change.

• Builds with Arduino 2.0.x (ESP-IDF 4.4) / 3.x and ESP-IDF 5.1 to 6.1 across ESP32, S3, C3, C5, C6, C61, H2 and P4.

Not tested

• ESP32-S2, C3, C6 and H2 hardware (the changed code has no chip-specific branch).