Inflate Covariance in CKF

[beomki-yeo]

While working on #844, I found that I didn't add covariance inflation in the CKF. This PR adds it for CPU CUDA and SYCL implementations.

It seems thrust does not support the lambda expression by default so I defined a struct for covariance inflation to make a functor.
However, SYCL does not seem to work with a functor. For example, when I use the following code:

    covariance_inflator cov_inf(config.covariance_inflation_factor);
    oneapi::dpl::for_each(policy, in_params_device.begin(),
                          in_params_device.end(), cov_inf);

I got the following error:

/opt/intel/oneapi/dpl/2022.7/include/oneapi/dpl/pstl/hetero/../hetero/dpcpp/parallel_backend_sycl.h:242:17: error: no matching function for call to object of type 'const oneapi::dpl::unseq_backend::walk_n<oneapi::dpl::execution::device_policy<> &, traccc::covariance_inflator>'

If I use the lambda expression for SYCL like in this PR, the compilation takes forever... any ideas?

[sonarqubecloud]

Quality Gate passed

Issues
1 New issue
0 Accepted issues

Measures
0 Security Hotspots
0.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

[beomki-yeo]

The increased covariance makes CKF pick up the measurements very far from the track. ($$\chi^2 \propto distance^2/covariance$$). Eventually CKF will produce lots of fake tracks and increase the wall time significantly 🤔
I will leave this PR draft until I find a solution

[stephenswat]

Would it help if we found measurements in a sorted way, i.e. we select the $N$ best measurement (ranked by $\chi^2$) instead of what we do now which is selecting $M$ random measurements within a $\chi^2$ cut; with $N < M$?

[stephenswat]

Note this is something I have discussed as being useful with @paradajzblond as well.

[beomki-yeo]

We already have max_num_branches_per_surface defined in finding_config.hpp, so it won't be very difficult. Maybe we should have implemented that already?

However, I am afraid this approach still does not solve the issue from this PR. CKF will still make branches as many as $N_{max}$.