Profile the gpt-2 module and the tortoise-tts gpt-2 module and try to improve the gpt-2 module's performance

[balisujohn]

This task is blocked by the gpt-2 forward pass test being added since this could introduce regressions. #5

The task is as follows:

measure the runtime of the autoregressive model all the way from inputs to the full sequence of tokens and last layer latents being generated (as checked by the test in #5 ), and the time taken for the corresponding batch of 4 token sequences and final layer latents in tortoise-tts.

Then, try improving the efficiency of the tortoise.cpp forward pass. Some suggestions are as follows:

try removing seemingly redundant ops
change ops to in place where possible

Feel free to ask questions here or in the discord.