handle transient and permanent cuda failures differently

[totaam]

Some of these errors are worth retrying - others are not.

For example, this one is unlikely to yield better results later:

RuntimeError: creating decoder returned error: 'NO_DEVICE'
2024-09-04 16:46:25,252 Error drawing on window 2
Traceback (most recent call last):
  File "/usr/lib64/python3.11/site-packages/xpra/client/mixins/window_manager.py", line 1592, in _do_draw
    window.draw_region(x, y, width, height, coding, data, rowstride,
  File "/usr/lib64/python3.11/site-packages/xpra/client/gui/client_window_base.py", line 838, in draw_region
    backing.draw_region(x, y, width, height, coding, img_data, rowstride, options, callbacks)
  File "/usr/lib64/python3.11/site-packages/xpra/client/gui/window_backing_base.py", line 923, in draw_region
    self.paint_with_video_decoder(VIDEO_DECODERS.get(coding),
  File "/usr/lib64/python3.11/site-packages/xpra/client/gui/window_backing_base.py", line 800, in paint_with_video_decoder
    vd.init_context(coding, enc_width, enc_height, input_colorspace)
  File "xpra/codecs/nvidia/nvdec/decoder.pyx", line 507, in xpra.codecs.nvidia.nvdec.decoder.Decoder.init_context
    self.init_decoder()
  File "xpra/codecs/nvidia/nvdec/decoder.pyx", line 593, in xpra.codecs.nvidia.nvdec.decoder.Decoder.init_decoder
    cudacheck(r, "creating decoder returned error")
  File "/usr/lib64/python3.11/site-packages/xpra/codecs/nvidia/cuda_errors.py", line 17, in cudacheck
    raise RuntimeError(msg)

[totaam]

Improved somewhat in 4ce1d9b + a61ea33

We need to remove the decoder spec completely if the error is not transient..
Problem is that some other errors will also show up as RuntimeErrors..

[totaam]

Forgot to tag this ticket: 97d1ac3, a071047, 9c2953f

[totaam]

I think that the commits above improve things quite a bit, but the new problem this creates is that we may now end up disabling video decoders that were previously used to expose encodings to the server!

So we may need to tell the server that the list of encodings has changed.
Perhaps overload the encoding packet with an optional list of encodings? Or full csc modes for each one?

[totaam]

This will do for now.