Stop LocalDataCluster reads from re-applying quirk value conversions

[rasmusbe]

Proposed change

LocalDataCluster.read_attributes_raw serves values straight from the attribute
cache, and zigpy feeds every successful read result back through
_update_attribute (via _legacy_apply_quirk_attribute_update in
Cluster.read_attributes). On a cluster that converts values in
_update_attribute, that converts an already converted value a second time and
stores the result, so every read corrupts the value further.

LocalIlluminanceMeasurementCluster applies 10000 * log10(v) + 1, so a reading
of 100 lx walks towards the fixed point of that function on each read:

100 -> 20001 -> 43011 -> 46336 -> 46660

20001 is the correct stored value for 100 lx. 46660 decodes to roughly
46 300 lx — direct sunlight. I observed this on an Aqara Curtain Driver E1 after
four reads of measured_value in the "Manage Zigbee device" UI, but nothing
about it is specific to that device: it affects every quirk pairing
LocalDataCluster with a converting _update_attribute, which includes all
Xiaomi quirks using LocalIlluminanceMeasurementCluster.

The cache is the source of truth for a local cluster, so an echo of a value it
just returned carries no new information. This records the attribute ids handed
back by read_attributes_raw and drops the matching _update_attribute call.

Only the attribute id is compared, not the value. A subclass overriding
_update_attribute sits ahead of LocalDataCluster in the MRO and has already
transformed the value by the time it gets there — comparing values works for
LocalIlluminanceMeasurementCluster but not for
IlluminanceMeasurementClusterP1, which overrides _update_attribute itself.

Additional information

Reproduction: on any device using LocalIlluminanceMeasurementCluster, read
measured_value repeatedly on the local cluster and watch the value climb.

Two regression tests, both verified to fail without the change:

• test_local_data_cluster_read_does_not_reconvert covers the pattern generally
  with a synthetic converting local cluster
• test_xiaomi_local_illuminance_survives_repeated_reads covers the real
  LocalIlluminanceMeasurementCluster via LumiMotionAC02

Side effects worth flagging for review: reads of _CONSTANT_ATTRIBUTES and
_DEFAULT_VALUES no longer populate the attribute cache as a side effect of the
echo. The read itself still returns those values, and get() still falls back to
_DEFAULT_VALUES, so I could not find behaviour that depends on the old
caching — but it is the one observable difference beyond the fix itself.

There is a narrow window between read_attributes_raw and the echo in which a
genuine update to the same attribute id would be dropped. read_attributes_raw
on a LocalDataCluster performs no I/O, and updates to a local cluster come
from quirk code running synchronously inside a report handler, so I do not
believe this is reachable in practice.

No separate issue was filed; this PR is the report.

Device diagnostics

Not applicable — this is a fix in LocalDataCluster, not a new or changed quirk.
The behaviour is covered by the two tests above rather than by diagnostics data.

Checklist

• The changes are tested and work correctly
• pre-commit checks pass / the code has been formatted using Black
• Tests have been added to verify that the new code works
• Device diagnostics data has been attached (not applicable, see above)

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 92.60%. Comparing base (6a3822c) to head (12ba5b3).

Additional details and impacted files

@@           Coverage Diff           @@
##              dev    #5287   +/-   ##
=======================================
  Coverage   92.60%   92.60%           
=======================================
  Files         424      424           
  Lines       14667    14676    +9     
=======================================
+ Hits        13582    13591    +9     
  Misses       1085     1085           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.