Define TORCHTEXT_API macro for visibility control

[mthrok]

Context:

TorchText uses dual-binding (PyBind11 and TorchBind) to make custom operations available in Python.
The both binding eventually calls the same implementation contained in libtorchtext.so.
The ones bound via PyBind11 (the ones in torchtext._torchtext) calls into libtorchtext.so.

This means that libtorchtext.so has to make the symbols (APIs) used by torchtext._torchtext visible.

However, the default visibility of symbols in shared libraries are different in Windows.
On Windows all the symbols are by default hidden. To work around this, we use CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS to expose all the symbols.
There is an upper limit of visible symbols one library fine can contain, and this can be problematic in the future.
(Although it is unlikely that torchtext will hit the limit, unless it introduces custom CUDA kernels.)

A better approach is to selectively mark the symbols that should be visible as visible.

Summary of the change set

This commit introduces TORCHTEXT_API macro which annotates functions with proper visibility.
The core logic was taken from https://github.com/pytorch/pytorch/blob/bcc02769bef1d7b89bec724223284958b7c5b564/c10/macros/Export.h

The behavior is as follow;

For non-Windows: It is always __attribute__((__visibility__("default")))
For Windows:
If the header is included from the compilation unit of libtorchtext, then it resolves to __declspec(dllexport). otherwise it resolves to __declspec(dllimport).

This allows to remove CMAKE_WINDOWS_EXPORT_ALL_SYMBOLS.

[Nayef211]

LGTM! Thanks @mthrok for introducing the change to add correct visibility for Windows systems.

[mthrok]

Test for fb internal https://www.internalfb.com/diff/D37475369