[ESSREDUCE] refactor: import ess.reduce submodules lazily

[SimonHeybrock]

ess/reduce/__init__.py eagerly imports nexus, normalization, polarization, uncertainty and unwrap, so importing any part of the package pays for all of it. Because importing a submodule always executes the parent __init__, an importer that wants one piece has no way to opt out.

lazy.attach defers each submodule to first attribute access, the pattern scippneutron already uses in its own __init__. streaming is now listed alongside the others — it was importable but missing from __all__.

Effect

Measured on this branch, python -c "<import>", best of three:

	eager	lazy
import ess.reduce	665 ms, 177 MB, 1342 modules	39 ms, 17 MB, 154 modules
from ess.reduce import streaming	715 ms, 177 MB, 1343 modules	244 ms, 60 MB, 676 modules
from ess.reduce import unwrap	737 ms, 176 MB	720 ms, 176 MB

Someone who wants unwrap pays what they always did. Someone who wants streaming no longer pays for scippnexus, scippneutron and — when it is installed — numba, none of which streaming touches.

Downstream savings depend on what else the application imports. ESSlivedata's dashboard imports ess.reduce.streaming for the accumulators in its Kafka message adapter and never interpolates a lookup table; its total import time is unchanged, because it loads panel, bokeh, holoviews and scipp regardless, but its resident set drops from 287 to 208 MB. That case is what prompted this: scipp/esslivedata#1253 makes numba a hard dependency so backend services stop silently falling back to the SciPy interpolator, which would otherwise put ~68 MB of numba/llvmlite into every dashboard process for a module it never calls.

Relation to #707

#707 moves the numba import inside _get_interpolator_class, which removes the forced numba import from this path independently of this PR. The two overlap but neither replaces the other. For from ess.reduce import streaming, best of seven:

	time	maxrss	modules
today	734 ms	177 MB	1343
with #707	563 ms	110 MB	1078
this PR	211 ms	60 MB	676

The #707 row is approximated by making numba unimportable, so that path falls back to SciPy; #707 itself imports neither interpolator at module level, which would make its row marginally better still. Whichever lands first, the other keeps most of its effect: #707 accounts for the numba share, this PR for the nexus -> scippnexus/scippneutron tree.

Notes

• __init__.pyi keeps the submodules visible to type checkers, as lazy.attach alone hides them.
• lazy_loader is added to dependencies. It was already installed everywhere via scippneutron, but it is now a direct import.
• pixi.lock is untouched: lazy_loader 0.5 already resolves into every environment, and CI runs frozen: true. Regenerate if your convention requires it regardless.
• Behaviour is unchanged for callers: dir(), attribute access on ess.reduce, and from ess.reduce import X all work as before. Deferring is safe because these submodule __init__s are pure re-export aggregations with no import-time registration.

Test plan

• pytest tests for essreduce: 912 passed
• ESSlivedata's suite against this branch: 4768 passed
• ruff check / ruff format clean on the changed files