Optimization of Matrix struct

[irakliyk]

Currently, Matrix struct stores trace and polynomial evaluation data in a column-major order. This allows us to run LDE for each column independently, but presents challenges during constraint evaluation where we need to process two sequential rows. Thus, we spend some time reading rows from the trace by copying values from each column into an EvaluationFrame struct.

A better approach could be to store data in a Matrix in a row-major order. This has the following advantages:

1. EvaluationFrame can just store pointers to rows in the trace - no need to move any data around.
2. Running LDE and FFT over multiple columns in parallel will actually be faster - probably up to 20% faster than doing it column by column as we do now.

[irakliyk]

Closed by #124, #142, #163, and #168.