Memory use in BSE runs

[sangallidavide]

When computing the BSE kernel, the yambo code allocates in meory, at the same time:

1. the full BSE matrix
2. the wave-functions
3. the oscillators for exchange or correlation

This requires a lot of memory. On top of that, sometimes the full BSE matrix is also duplicated without distribution in memory.
See possible solution in issue #113

Different solutions are possible to limit the memory

1. the BSE matrix can be loaded in memory block by block at building time if I/O is on
   (see branch https://github.com/sangallidavide/yambo/tree/tech-isolv-bse)
   Only later, when the solver is called, the full BSE mat is loaded in memory, while reading from disk.

2. WFs can be loaded in memory k by k with a proper implementation. See also (discussion with @attacc):
   2.a - k_extrema_only option in https://github.com/yambo-code/yambo/blob/tech-master/src/wf_and_fft/WF_load.F
   2.b - implementation in https://github.com/sangallidavide/yambo/tree/feature-shifted-grid

3. oscillators can be loaded independently in exchange and correlation (see changes introduced with merge Tech bse #98 )