todo

1. remove Link in the setupPath file
2. handle the find nothing problem in preprocess
3. handle projection size larger than 28888

check the wavelet, it is very suspicious

Noticed that:

1. continuation.
2. the direct implementation continue decreasing the cost function, while
the fourier transform implementation stack at certain point, although the
two share the same path before the point.
The RSE can reach unblievable precision. 4e-5; for less than 1000 iterations

The line search for FISTA speed up the convergence in terms of number of outer
iterations.

FISTA_NNL1 and FISTA_ADMM_NNL1 are almost equivalent, which means with/without
line search, their convergence curve are the same.

Restarting in FISTA is not better.

todo for denoise_bound_mod
     use more agressive step size
     use C implementation
     use a better mask implementation

  make solver to support complex number

write a c++ version of the algorithm
  multi threads?
  gpu?

!!!make sure that bhc works!