Feature/eckit sparsematrix allocators reorganization, new in-place allocator

[pmaciel]

This add a new "allocator" to the SparseMatrix, which allows building one with direct, non-owned access to a set of Index*, Index* and Scalar*.

It allows low-level control of the data structures, benefiting the following:

• to provide views, hence promoting parallelism
• efficient handling of very large cases, by exploiting the internal structure (per-row assembly, non-linear corrections, etc.)
• creation of utilities that observe and manipulate the structure directly (reordering, manipulation of weights, etc.)

There's no urgency

[codecov-commenter]

Codecov Report

Attention: Patch coverage is 66.66667% with 23 lines in your changes missing coverage. Please review.

Project coverage is 63.75%. Comparing base (9d2807d) to head (e59c541).

Files with missing lines	Patch %	Lines
src/eckit/linalg/allocator/BufferAllocator.cc	0.00%	11 Missing ⚠️
src/eckit/linalg/allocator/InPlaceAllocator.cc	76.19%	5 Missing ⚠️
src/eckit/linalg/allocator/StandardAllocator.cc	68.75%	5 Missing ⚠️
src/eckit/linalg/SparseMatrix.cc	71.42%	2 Missing ⚠️

Additional details and impacted files

@@           Coverage Diff            @@
##           develop     #148   +/-   ##
========================================
  Coverage    63.74%   63.75%           
========================================
  Files         1066     1069    +3     
  Lines        55360    55397   +37     
  Branches      4086     4086           
========================================
+ Hits         35289    35317   +28     
- Misses       20071    20080    +9     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

---

🚨 Try these New Features:

• Flaky Tests Detection - Detect and resolve failed and flaky tests

[github-actions]

Private downstream CI failed.
Workflow name: private-downstream-ci-hpc
View the logs at https://github.com/ecmwf-actions/private-downstream-ci/actions/runs/11954033047.

[pmaciel]

Say a few more days before merge? I have changes depending on this.

[wdeconinck]

I think this is great! A few points.

minor point: I would change to something like "NonOwningAllocator" or "WrappingAllocator"

more significant point:
When the SparseMatrix is created this way, wrapping existing data, the user of this class should not be allowed to call several methods:

• SparseMatrix::reserve(rows, cols, nnz)
• SparseMatrix::load()
• SparseMatrix::transpose()
• SparseMatrix::setIdentity()
• SparseMatrix::prune()

In fact anything that is attempting to change the shape should be forbidden.

One option is to add a boolean flag to SparseMatrix::Shape be set by the Allocator which can be checked if changes to shape are allowed or not. Open to other suggestions.

[wdeconinck]

I believe a reinterpret_cast or const_cast or any other casting should not be necessary.