Apply Coil Compression on a subset of the kdata for efficiency

[aTrotier]

The current implementation is svd on the whole kdata which is especially long.

Two implementations are available in gadgetron :

1. PCACoilGadget
2. GenericReconEigenChannelGadget

The first one takes the 100 first profiles (keep only the central part of the readout (16 points)) did some processing and then perform the PCA. The PCA is then applied to the remaining profiles.

It is efficient for online reconstruction because the processing of all the profile is then done on the fly. But for 3D acquisition, the PCA is most of the time performed on 100 profiles acquired on the outer part of the k-space.

2. I did not really take a look at the second implementation but it seems it is applied after buffering all the profiles and on the calibration / central part of the k-space.

---

I don't have a perfect answer to that problem :
Ideally, we should perform the SVD on the central part of kspace, do we have an efficient way to find the kdata correspoding to it ? (without creating the kspace with kDataCart())

Maybe we should also add an option to do it directly to the raw object which will select all the lines flags

"ACQ_IS_PARALLEL_CALIBRATION"              => 20,
"ACQ_IS_PARALLEL_CALIBRATION_AND_IMAGING"  => 21

in that case, a reconstruction from ISMRMRD files will be something like the gadgets in Gadgetron :

1. raw = RawAcquisitionData(ISMRMRDFile("path.mrd")
2. remove_oversampling!(raw)
3. software_coil_compression!(raw,6)
4. convert to acq
5. perform reconstruction