gomodfs, store/remotestore: fetch less over the network in WinFsp mode

[bradfitz]

Two changes that cut HTTP requests from a WinFsp mount to its remote
store:

Cache metadata files (.info, .mod and .ziphash) in remotestore.Store.
They never change for a module version, but every stat and open of one
fetched it from the server again, and the go command reads hundreds of
them per invocation, mostly one at a time while loading the module
graph. Over 8 cached go commands, that was about 13,600 requests at
about 1ms each. Like modmaps, they're cached for the life of the Store,
with concurrent fetches merged. Misses aren't cached.

Fetch a file's contents on its first read rather than when it's opened.
Windows opens files for many things that don't read them, including
every os.Stat, and each open downloaded the whole file. Opens now get
the size and type from the store's (cached) modmap, so an HTTP error
fails the first read instead of the open.

In a Windows VM with the module cache and Tailscale Go toolchain served
from a remote store, fully cached go commands (median of 30, p < 1e-10):

go build tailscale.com/cmd/tailscale/cli: 4.90s -> 2.21s
go run github.com/tc-hib/go-winres:       2.59s -> 1.58s

and the first command after mounting goes from 5.73s to 3.10s and from
2.50s to 1.26s. (On NTFS, both take under 0.3s.) "go mod verify" of all
75 modules of the build still passes over the mount.

Updates tailscale/corp#24037