perf: gate devtools in prod and lazy-load leaflet to reduce initial js

[cc-cruz]

rebased onto upstream/master; minimal diff

disclaimer: couldn’t find a matching web bounty card for this exact change. i wanted to get my feet wet with a small, safe performance improvement before picking up a bigger approved bounty. if appropriate, please attach this pr to the relevant web bounty/card from the board.

summary

• reduce first‑load javascript and keep within the bundle budget
• load solid devtools only in development
• lazy‑load leaflet in device location so map code downloads only when a map is shown

changes

• src/index.tsx: conditionally import solid-devtools when import.meta.env.dev is true
• vite.config.ts: register solid-devtools/vite only in development
• src/components/devicelocation.tsx: replace top‑level import 'leaflet' with dynamic import in onmount; types are import‑only

metrics (local prod build)

• main entry chunk (gzip): 68.90 kb → 26.47 kb (~−42.43 kb, ~62%)
• leaflet chunk (gzip): ~42.51 kb split and loaded on demand
• total compressed size: ~268 kb (under 270 kb)

lighthouse (local prod preview http://localhost:5176, experimental preset)

• fcp ~2.1 s, lcp ~2.5 s, speed index ~2.1 s

verification

• bun run build
• bun run serve -- --port 5176
• bun run bundle-size
• bun run check
• bun run test -- --run
  manual
• non‑map pages start with smaller js download
• navigating to a map screen loads the leaflet chunk on demand; map renders normally

risks

• map uses dynamic import; failures are localized to the map screen and easy to revert

follow‑ups (optional)

• defer map render until in‑viewport (intersectionobserver)
• add a ci gate that fails on >5% bundle regression
• prefetch map/video chunks on hover/intent

[github-actions]

This branch is behind commaai/master. The line count diff bot is disabled.

[greatgitsby]

cool refactor! in my experience, comma typically prefers smaller, focused PRs. i'd split this into two since they are separate improvements