Feature/mpu6050 drdy interrupt - #19601

Summary

The MPU6050 driver only implemented fetch(), so samples were read on
demand: timestamped when the application asked rather than when the device
measured, with the scheduler's jitter baked in, and with accelerometer and
gyroscope coming from two separate I2C reads so the two topics were never
sampled at the same instant. The part has an INT pin and a data ready
interrupt; nothing used them.

Two commits:

1. Push mode behind CONFIG_SENSORS_MPU6050_INT, plus a DLPF fix. The
   board supplies mpu6050_config_s::attach to wire the INT pin; the
   handler timestamps and defers to HPWORK, and the worker reads the device
   once and pushes both topics from that single sample. The I2C read cannot
   run in the interrupt, hence the work queue. The mode is chosen at build
   time, so fetch() is simply left out of the ops table and out of the
   build when the option is set: an instance uses one model or the other,
   never the mixture that made poll() unusable on l3gd20 (drivers/sensors/sensor: fix poll()/read() for fetch()-only sensors #19596). A board
   that enables the option without supplying attach is misconfigured, so
   registration fails with -EINVAL rather than silently falling back.

   The same commit sets CONFIG so the DLPF is on. The sample rate is the
   gyroscope output rate divided by 1 + SMPLRT_DIV, and that output is
   1 kHz only while the DLPF is enabled. CONFIG was left at its 0 reset
   value, which disables the DLPF and puts the output at 8 kHz, so the
   divider of 9 gave 800 Hz rather than the intended 100 Hz. fetch() hid
   this because the application set the pace and simply read the most recent
   sample; the interrupt made the real rate visible.

2. Update the driver documentation for the new registration signature
   and the two acquisition modes.

Impact

Boards that wire the INT pin get samples timestamped at acquisition, both
topics sampled from the same read, and one I2C transaction per sample
instead of one per read(). Boards that do not wire it are unaffected:
the option defaults to n and the fetch() path is unchanged. The DLPF
fix applies to both modes and corrects the configured rate to the 100 Hz
the code already documented, additionally giving the anti-alias filtering
that a 100 Hz sampler should have.

Testing

Host: Ubuntu 24.04.3 LTS. checkpatch.sh (style + -m commit messages)
clean on all four commits.

Compiles clean in both modes — sim (x86_64) with
CONFIG_SENSORS_MPU6050_INT on and off, and xtensa-esp-elf-gcc 14.2.0 for
the hardware runs below. The refactor was re-verified on hardware after the
fact: same 101 Hz and no duplicates, so it is behaviour neutral as claimed.

On hardware — ESP32-S3-DevKitC + MPU6050 (GY-521), I2C0 on
SDA GPIO5 / SCL GPIO4, INT on GPIO6. The board glue providing attach()
for that target is not part of this PR.

Sample rate, measured over 100 samples of uorb_listener sensor_accel0,
before and after the DLPF commit:

The duplicates before the fix are the worker being re-triggered faster than
the device produced new data. After it, every interrupt yields a fresh
sample at the intended rate.

Both topics from one interrupt, note the shared timestamp:

nsh> uorb_listener -n 6 sensor_accel0,sensor_gyro0
Monitor objects num:2
object_name:sensor_gyro, object_instance:0
object_name:sensor_accel, object_instance:0
sensor_gyro(now:72560000):timestamp:72560000,x:2.424675,y:0.108184,z:0.039836,temperature:22.394705
sensor_accel(now:72560000):timestamp:72560000,x:0.790087,y:1.301249,z:10.234015,temperature:22.394705
sensor_gyro(now:72570000):timestamp:72570000,x:-0.380109,y:0.310962,z:0.613530,temperature:23.668234
sensor_accel(now:72570000):timestamp:72570000,x:0.776918,y:1.278504,z:10.190920,temperature:23.668234
sensor_gyro(now:72580000):timestamp:72580000,x:-0.024781,y:0.089131,z:0.303501,temperature:23.682940
sensor_accel(now:72580000):timestamp:72580000,x:0.785298,y:1.274912,z:10.144233,temperature:23.682940
Object name:sensor_gyro0, received:3
Object name:sensor_accel0, received:3
Total number of received Message:6/6

Teardown: 8 rounds of uorb_listener -r 5 sensor_accel0 & followed by
SIGINT mid-poll. ps afterward shows no leaked task and an idle hpwork
thread, and uorb_listener -n 3 sensor_accel0 right after still completes
3/3 — the interrupt is disabled when the last subscriber leaves and
re-enabled when one returns.