feat: pre transform direct imports before requests hit the server

[yyx990803]

Currently Vite only transforms a file (i.e. go through the plugin pipeline) when the file is requested. This ensures only files actually used on the current page are transformed, however it also creates a problem with deeply nested import chains:

A --> B --> C

B is only requested when A is evaluated in the browser (triggering the import request to B). Similarly C is only requested after B is evaluated in the browser. The problem becomes more obvious when there are many modules being loaded in parallel - since the browser typically can only handle 6 parallel requests so it will hold off some requests until previously pending requests have resolved (HTTP/2 requires HTTPS which in fact make things slower). This causes requests to hit the server at a delayed rate so the server is in fact under-utilized and waiting for the browser. The deeper the import chain and the more modules are loaded in parallel, the more obvious the problem is.

Since Vite already performs import analysis when transforming each module, we can in fact start transforming all direct imports of a module before even sending back the response. This essentially means Vite will be eagerly compiling the non-lazy-imported sub graph of the application without waiting on network requests. When the requests come in, they can pick up the already initiated transform Promise (can be pending or already resolved).

I did some synthetic benchmarking here:

	D2/C5/T31	D3/C7/T400	D4/C4/T341	C1/C100/T101	D4/C8/T4681
Base	157ms	778ms	727ms	282ms	8092ms
Post Optimization	112ms (-28.6%)	321ms (-58.7%)	305ms (-58%)	142ms (-49.6%)	2414ms (-70.2%)

D = layers of nested imports / C = parallel imports per file / T = total number of modules loaded

---

Tests are passing but this may have subtle implications in real world cases, so we should put it in a beta and test it in some projects to make sure it doesn't break anything.