Quite a lot going on in this PR:
 * Updates to support recent emscripten, switching to WASM which is now the default
 * Use emscripten's MODULARIZE option rather than wrapping it ourself, since doing
   so in pre-post js doesn't work anymore.
 * Most changes are moving the emscripten runtime functions to top-level
   calls rather than in the Module object.
 * Get rid of duplicated NULL_BYTE_PADDING_LENGTH
 * Fix ciphertext_length used without being declared
 * Fix things that caused the closure compiler to error, eg. using
   OLM_OPTIONS without a declaration.
 * Wait until module is inited to do OLM_ERROR = olm_error()

The main BREAKING CHANGE here is that the module now needs to initialise
asyncronously (because it has to load the wasm file). require()ing olm
now gives a function which needs to be called to create an instance.
The resulting object has a promise-like then() method that can be used
to detect when the module is ready. (We could use MODULARIZE_INSTANCE
to return the module directly as before, rather than the function,
but then we don't get the .then() method).

olm_import_inbound_group_session, which reads the format written by
olm_export_inbound_group_session to initialise a group session.

A pair of functions which allow you to export the megolm keys for an inbound
group session, so that an application can save/restore them.

Avoid leaving copies of the plaintext sitting around in the emscripten heap.

Messages can be very large, so we don't really want to allocate them on the
stack. Switch to using the heap for them, and try to clean up some of the
string handling while we're at it.

Applications can use the index to detect replays of the same message.

Fix a couple of places where we were using the wrong method to get the last
error.

Repeat the fix from b10f90d for megolm messages.

It turns out that the 'length' argument to 'Pointer_stringify' doesn't work if
the input includes characters >= 128.

Rather than try to figure out which methods can return UTF-8, and which always
return plain ascii, replace all uses of Pointer_stringify with a 'length'
argument with the version that expects a NULL-terminated input, and extend the
buffer by a byte to allow space for a null-terminator.

In the case of decrypt, we need to add the null ourself.

Fixes https://github.com/vector-im/vector-web/issues/2078.

message.length counts codepoints; we need bytes.