reduce circuit fix and permutation data

[mooreland]

currently pk circuit data quite large and >5G with k23
fixed data(256bit) occupy >80%
permutation data occupy ~15%

this pr just optimize the datas:
fixed data: only store valid bit
permutation data: only store permuted cell

after optimization, pk data shrink to ~200M

meanwhile, added parallel processing
store time reduce to 300ms from 1.3s
fetch time reduced to 2s from 2.3s

[lanbones]

I believe the compression of sparse matrix is the most important. I have concern about the zero reduction progress. It involves mont and unmont computing in from_repr()/to_repr(), and lots of memory allocation on vec![]. So my options are:

1. How much benefit without the zero reduction progress?
2. Can we use zlib instead of zero reduction progress to avoid from_repr()/to_repr()?
3. Is it possible to remove these small vec![]?

[mooreland]

1. How much benefit without the zero reduction progress?
   --actually, the fix zero reduction occupy >80% size in pk, the matrix only occupy ~15%
   we can calculate them, 13 permutation columns and 44 cell involved, for k23, 2^23 *13 *4=436M totally.
   for 17 fix columns, for k23, 2^23 *17 *32=4.5G, actually, most of fix value are 0 or 1(enable), there are quite much meaningless 0. if we keep fix unchanged, the pk size will have 4.5G more. and with zero reduction, it only have ~200M

2. Can we use zlib instead of zero reduction progress to avoid from_repr()/to_repr()?
   --maybe not so efficient than zere reduction, we can try.
3. Is it possible to remove these small vec![]?
   -- not so clear, what is means for small vec![]?