Add drift_g2l passmethod

[xresende]

I noticed we are not using t_in, t_out, r_in and r_out in drift and identity passmethods. This came as a surprise for me when I tried to generate orbit signatures of dipolar perturbations at the center of straight sections (for IDs). Sometimes it is useful to be able to do such a thing, like adding dipolar perturbations at ring locations with such elements (center of straight sections as I did, or at BPMs, for example...), but most of the time applying these global to local transformations and vice-versa are only useless calculations with additional time costs.

Another possibility is to add a simple passmethod that can be switched to whenever these transformations are in need for markers or drifts. I also added this possibility in this PR (pm_drift_g2l_pass).

Which one should we use ?

I measured the additional time cost of running g2l and l2gat all drifts and markers in our current SI model:

import time
import numpy as np
import pyaccel
from pymodels import si

MODEL = si.create_accelerator()

def tracking(nrparts, nrturns):
    particles = np.zeros((6, nrparts))
    t0 = time.time()
    pyaccel.tracking.ring_pass(MODEL, particles, nrturns)
    t1 = time.time()
    print(f'time {t1-t0:.3f} s')

# before PR
%timeit tracking(1, 5000)
time 6.867 s
time 6.912 s
time 6.826 s
time 6.789 s
time 6.775 s
time 6.725 s
time 6.857 s
time 6.714 s
6.8 s ± 65.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

# after PR
%timeit tracking(1, 5000)
time 6.971 s
time 7.004 s
time 7.065 s
time 7.062 s
time 7.124 s
time 7.639 s
time 7.684 s
time 7.528 s
7.3 s ± 279 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

in this final version g2lwas removed from identity_passand drift_pass. a new drift_g2l was added to allow for phase-space space parameters shift and rotation at elements entrances and exits.

[fernandohds564]

@xresende, I don't understand how this PR can change the execution times of the code in the description, because it only adds a new pass_method, inexistent in the default sirius model. Does it increase the execution times just because of the existence of the new pass_method?

Another point is that I think this new pass_method should not be a drift with the global2local transformations taken into account. I think it could just be a pure transformation of the beam coordinates, with a more appropriate name. What do you think?

[xresende]

@xresende, I don't understand how this PR can change the execution times of the code in the description, because it only adds a new pass_method, inexistent in the default sirius model. Does it increase the execution times just because of the existence of the new pass_method?

Another point is that I think this new pass_method should not be a drift with the global2local transformations taken into account. I think it could just be a pure transformation of the beam coordinates, with a more appropriate name. What do you think?

@fernandohds564 , as I discussed with you outside github, this PR is a mixture of two possible options: adding g2l back-and-forth transformations in identity and drift passmethods (implemented here!) or adding a new passmethod drift_g2l (also implemented here). I opened this PR so that you guys can answer my posed question: which one is better to keep ?

the time estimates were evaluated using standard drift and identify passmethods with additional g2l and l2g conversions, (as is coded here!), so they do influence the tracking times.... I am confused with you questioning...

regarding you second point: I think we may want to add a coordinate shift or/and rotation in elements that now are drifts in our models. having the possibility of doing this in drifts is the most practical way: avoid changing the model just for adding such transformations in drifts but only changing the passmethod index of the element. also, these drift_g2l pass methods can be used for markers replacement too since de default length=0 is used and generate and identity passmethod in practice. But if computation cost of using a drift(length=0) for markers is to be avoided, we can also add a second new passmethod: one for drift with g2l_transf, another for marker with g2l_transf...

the name drift_g2l i did not like either, but opened the PR anyways so that you guys can make suggestions. maybe we can change to another thing. g2ltransf (g2ltransf_drift and g2ltransf_marker in case of separate new passmethods) or similar?