Update nvda_dmp to possibly address an autoread hang

[codeofdusk]

Link to issue number:

Possible resolution for #11992.

Summary of the issue:

Currently, nvda_dmp sends the number of Unicode characters in the response header, then sends utf-8 encoded data. NVDA expects the header to contain the number of bytes, resulting in the diff hanging when the number of characters is fewer than the number of bytes.

Description of how this pull request fixes the issue:

nvda_dmp now sends the number of bytes (not characters) in the response header.

Testing performed:

I can no longer reproduce the hang described in #11992 and #11639 (comment) (I previously thought it was an issue with locking until further testing revealed this bug).

Known issues with pull request:

• This PR definitely improves behaviour for me, but should be merged ASAP to allow for testing on master.

Change log entry:

None needed.

[codeofdusk]

Cc @feerrenrut, @Neurrone.

[feerrenrut]

From https://stackoverflow.com/a/54063928 I expected that at each "synchronization boundary" between NVDA and the DMP process there should be a call to flush, in both processes.

Eg if this code is sending some data and waiting for DMP to read it and respond, there should be a flush of std-out before the read. Otherwise you could end up with the DMP process waiting on extra data from NVDA, and NVDA waiting on data from DMP, therefore a deadlock.

[feerrenrut]

So for instance after line 80 DiffMatchPatch._proc.stdin.write(new)