feat(vfs) Try to unify file descriptor

[Hywan]

Unfortunately, this patch contains multiple features.

First off, it renames the types from host_fs and mem_fs so that
they use the same names. Apart from the module name, the types owned
by the module are the same, which simplifies the usage greatly.

Second, the code has been cleaned up a little bit. The more important
clean up is probably the renaming of MemKind to Node. It's OK to
do that here, because the code is not released yet.

Third, a new FileDescriptor type is introduced. It represents a
“generic” file descriptor, that a VirtualFile can hold. It contains
a value with the size of the larger size a file descriptor on any
platform, namely on Windows where a file descriptor is the size of a
pointer (compared to Unix where it's a i32). Then a bunch of
TryInto implementations are written to convert it to RawFd (on
Unix) or RawHandle (on Windows). I'm not confident it's the best
design for the moment, but it's better than before where Windows file
descriptors were impossible to represent. Note that we cannot make
FileDescriptor generic over the system for 2 reasons

1. We can activate multiple file systems at the same time, so the
   representation of a file descriptor can change between file system,

2. If we make FileDescriptor generic, then the size of dyn VirtualFile is unknown at compile-time, which makes impossible to
   use in wasmer-wasi, hence this design decision of taking the
   larger possible size for a file descriptor so that it fits all
   scenarios.

[syrusakbary]

I believe a strategy using unsafe-io will be better for the cross-platform abstraction.
https://docs.rs/unsafe-io/0.7.1/unsafe_io/

UnsafeFile, UnsafeSocket or UnsafeHandle.

Resource	Posix-ish type	Windows type		Platform-independent types
File/Pipe	RawFd	RawHandle		UnsafeHandle or UnsafeFile
Socket	RawFd	RawSocket		UnsafeHandle or UnsafeSocket
Any	RawFd	RawHandleOrSocket		UnsafeHandle

[Hywan]

Actually we cannot do that.

In our case, we cannot have multiple file descriptor implementations. Why? Because the entry point is VirtualFile, which has a get_(raw)_fd method. It must return a statically known value, otherwise the size of VirtualFile cannot be known at compile-time. So we cannot use a FileDescriptor implementation from host_fs or mem_fs, or even both depending on how wasmer-vfs is compiled.

This patch contains a little bit more clean up, but everything compiles gently, even with wasmer-wasi.

Calling about wasmer-wasi, from a VirtualFile, to get a RawFd, we now just need to do: virtual_file.get_fd().try_into().unwrap() and that's it. The addition compared to the initial code is try_into. We could use an Into implementation but I prefer to provide safe abstraction and not assuming that the file descriptor has been constructed correctly.

I've also tried but cancelled, to “connect”/“link” a FileDescriptor to a filesystem instance, but it actually makes VirtualFile really complex. The gain doesn't worth it I suppose for the moment.