perf: Defer heavy imports on Python 3.15 with lazy imports

[karangattu]

Stacked on #2519 — this PR targets add-python-3-15-support and will retarget to main automatically when #2519 merges.

What

Python 3.15 adoption follow-up from #2519:

• Lazy imports (PEP 810) on Python 3.15+. __lazy_modules__ declarations defer the modules that dominate import shiny but are only needed for specific features:

  • narwhals and the render modules: shiny.render._data_frame, shiny.render._data_frame_utils._selection, shiny.render._data_frame_utils._types, shiny.render._render
  • shinychat: shiny.ui._chat, shiny.ui._markdown_stream (mirrored in shiny.express.ui)
  • the CLI/server stack (shiny._main): uvicorn, click, questionary, prompt_toolkit, watchfiles

  Every deferred module reifies transparently on first use. __lazy_modules__ is ignored on Python 3.11–3.14, so behavior there is unchanged.

• Explicit UTF-8 for text file I/O (theme CSS/SCSS, shiny create templates, bundled Agent Skills, pytest generator data). These previously used the locale encoding, so non-ASCII content could mis-handle on Python 3.14 and older (e.g. cp1252 locales on Windows); 3.15 already defaults to UTF-8.

Measured effect

scripts/benchmark_startup.py (included in this PR): median wall-clock of fresh subprocesses, so the numbers include interpreter startup, on a warm filesystem cache — Apple M2 Pro (macOS), shown with min in parentheses.

scenario	Python 3.12.15	Python 3.15.0, before	Python 3.15.0, this PR
bare interpreter	42 ms	37 ms	37 ms
import shiny	686 ms (666)	308 ms (276)	143 ms (122) — −54%
from shiny import App, ui	692 ms (653)	303 ms (289)	142 ms (129) — −53%
app boot (App(ui.page_fluid(ui.h1("hi")), None))	769 ms (703)	318 ms (283)	155 ms (132) — −51%

The Python 3.12 column is unaffected by this PR (the declarations are inert before 3.15); it's included for the cross-version comparison. Reproduce with:

uv run python scripts/benchmark_startup.py --runs 12

Notes and limitations

• Reification moves import errors and import-time side effects for the deferred modules to first use (the traceback still points at both sites, so debugging is unaffected).
• Deferral only helps code paths that never touch those features; an app that renders data frames or uses chat pays the same cost, just later.
• Other 3.15-only features (sentinel, frozendict, PEP 798 unpacking in comprehensions, TypeForm) were deliberately not adopted: they would need per-version gating and don't have a current use in this codebase.

App start time (shiny run)

Time from launching python -m shiny run app.py --port N to the first HTTP 200 on /, for a minimal app (ui.input_text + @render.text). Median of 12 runs after 2 warmups, fresh subprocess per run, warm caches, Apple M2 Pro (macOS), same wheel for the two 3.15 environments (3.15 base = add-python-3-15-support tip).

environment	min	median
Python 3.12.15	321 ms	345 ms
Python 3.15.0, before this PR	325 ms	367 ms
Python 3.15.0 + this PR	232 ms	280 ms

On the full shiny run path, shiny._main (uvicorn/click/questionary) is needed by the CLI itself, so the saving comes from shinychat and the narwhals render stack no longer being loaded before the server is ready: 3.15 alone is roughly even with 3.12 here, while 3.15 + this PR is ~20% faster to first response (min: −28% vs 3.15 base). After import app, none of shinychat, narwhals, questionary, prompt_toolkit, uvicorn, watchfiles, click are in sys.modules on 3.15 + this PR; all of them are on 3.15 base and on 3.12.