Improve perf for predict impact

Perf:

- [ ] Parallelize
- [ ] ~~Reuse the PathCalculator~~ Already done.
- [ ] Pass back road GJ once, and then just counts after that
- [x] Use a basic loading screen (switch to web workers?)
- [ ] Vec for counts, not HashMap
- [ ] Profiling
- [ ] ~~Dedupe requests upfront~~ Not enough duplication to be worthwhile
- [ ] Measure how long loading the demand model adds. Maybe async load it?
- [ ] Could speed up the Scotland demand data builder with rstar

这个 issue 是关于优化 'predict impact' 功能的。首先对当前实现进行性能分析，以确定瓶颈。检查 PathCalculator、需求模型加载，以及用于计数的 HashMap/Vec 等数据结构。确认是否适用 web workers 或异步加载。Scotland demand data builder 和 rstar 被提及为潜在的加速方案。

由索引模型根据 Issue 内容生成。