Accelerate collect() by switching to lazy loading

[mikekryjak]

This is another change that Claude came up with after #337. It rewrites xbout.load.collect() to use the new lazyload.lazy_open_boutdataset() from #336. It falls back on the original method if things break. It also makes minor improvements to the code, like not relying on the coordinates not being in a specific order.

Test results:

• xbout.collect() before change: 24.1s
• xbout.collect() after change: 0.12s (!!)
• boutdata.collect(): 0.20s

So we are now 200x faster than before, and also 40% faster than boutdata. I suppose the difference could be using Dask.

For completeness, the same dataset takes 2.5s to load in its entirety using the latest lazy loading.

This PR contains #336 and should be merged after.