fix(npu): alert on the temperature both interfaces already ask for

[ntdatt812]

The NPU plugin's temperature is never alerted, in either interface, at any reading.

update_views() does not extend the views the base class built — it replaces them:

self.views[i[self.get_key()]] = {'load': {}, 'freq': {}, 'mem': {}}

temperature is not in that dict, so views[npu_id]['temperature'] does not exist, and both readers fall back to no style:

• msg_curse asks get_views(item=..., key='temperature', option='decoration') (npu/__init__.py:249), and get_views returns the string 'DEFAULT' for a key it cannot find.
• plugin-npu.vue:35 calls getDecoration(npu.npu_id, 'temperature'), whose first line returns undefined when the view is missing, so no class is bound.

Both sides were written expecting a decoration that nothing produces. The reading itself is real: cards/intel.py:82 sets it from the driver (temp_value / 1000.0); AMD and Rockchip document it as not implemented and leave it None.

The sibling GPU plugin has had all three alerts — proc, mem, temperature — since it was written, so this is the NPU plugin missing one its own callers assume, not a feature neither has.

The change

Adds 'temperature': {} to the view dict and the alert beside the other three, in the GPU plugin's shape. A card that reports no reading keeps an empty view rather than being alerted on None, which get_alert would otherwise score through (None * 100) / 100.

conf/glances.conf gains the matching [npu] thresholds on the same 60/70/80 ladder [gpu] uses. That half is not cosmetic: with no limits configured, get_alert takes its not careful and not warning and not critical branch and returns 'DEFAULT', so the code change alone would be invisible.

I did not add mem_* thresholds to [npu], although update_views computes a mem alert that is unconfigured for the same reason. Leaving a stat unthresholded is a normal default across the plugins; temperature is different only because two call sites already read a style for it.

Verification

pytest tests/test_plugin_npu.py is 13 passed (new file); with test_plugin_gpu.py and test_plugin_quicklook.py, 56 passed / 9 skipped. ruff check and ruff format --check are clean.

Each half was checked by reverting it:

reverted	fails
'temperature' dropped from the view dict (the shape on develop)	all 9 plugin tests
the alert block removed, the view key kept	8 of 9 — the "driver reports None" test still passes, which is the discrimination it is for
the [npu] thresholds removed from conf/glances.conf	the 4 shipped-config tests only

The last row is why the config assertions exist at all: without them the code change could be reverted in conf/glances.conf alone and every behavioural test would still pass while the alert stayed invisible.

I have no NPU to hand, so the assertions drive update_views over a synthetic card rather than a real one.