Currently, when a controller is configured to use filtered mode, thermal
readings are valid only about 30% of the time.
Upon testing, it was noticed that lowering any of the interval settings
resulted in an improved rate of valid data. The same was observed when
decreasing the number of samples for each sensor (which also results in
quicker measurements).
Retrying the read with a timeout longer than the time it takes to
resample (about 344us with these settings and 4 sensors) also improves
the rate.
Lower all timing settings to the minimum, configure the filtering to
single sample, and poll the measurement register for at least one period
to improve the data validity on filtered mode. With these changes in
place, out of 100000 reads, a single one failed, ie 99.999% of the data
was valid.