0.10.6 — consistent entity-action error handling

No breaking changes. User-visible improvement: entity actions that fail
now show a clear error instead of either silently no-oping or crashing.

Closes a gap flagged in 0.10.4's round notes as "a cross-cutting decision
bigger than this pass's scope": button.py's 11 async_press methods,
select.py's 18 async_select_option methods, switch.py's 4
async_turn_on/async_turn_off methods, and number.py's 10
async_set_native_value methods had no handling at all for the
library's own SHCException/SHCConnectionError — a real API rejection
or SHC comms failure during a write propagated as a raw unhandled
exception instead of a clean, translated error. All 43 methods now follow
the same pattern already established in alarm_control_panel.py/
binary_sensor.py: catch (SHCException, SHCConnectionError), raise
HomeAssistantError with a shared per-platform translation key
(button_press_failed, select_option_failed, switch_action_failed,
number_set_failed — reusing the existing smoke_test_failed where the
action is literally a smoke test). Translated to all 30 languages.

quality_scale.yaml's action-exceptions rule was already marked done
but the claim was incomplete — it only covered two custom domain services,
not these 43 entity write methods. Corrected with an honest accounting of
what's covered now versus before.

climate.py's existing log-and-swallow behavior was deliberately not
changed: _async_apply_hvac_mode is a shared bool-returning helper used
by both async_set_hvac_mode and async_set_temperature, and the two
callers need to distinguish "mode write failed" from "mode is a no-op"
differently. Correctly disambiguating that so one caller can raise while
the other keeps its existing behavior is a real refactor of shared control
flow, not a mechanical wrap — risks a live behavior regression without
real-device verification, so it's tracked as a separate follow-up rather
than forced through blind.

15 new regression tests covering the error path (representative coverage
across all 4 files, not one per method — 43 near-identical error-path
tests would be redundant given they all exercise the same try/except
shape). 2988/2988 tests green, ruff/pylint/mypy/codespell clean, Gold
quality-scale gate still passes.