[FIX] unstable TorchBackend.sqrtm() around repeated eigenvalues

[samuelbx]

Types of changes

Implemented closed-form differential for TorchBackend.sqrtm (SPD matrix sqrt) with comparable time complexity (1 eigendecomposition) but stable around repeated eigenvalues

Motivation and context / Related issue

See issue #773

How has this been tested (if it applies)

• Minimal reproducing example from Unstable TorchBackend.sqrtm() around repeated eigenvalues #773 outputs OK
• Previous tests on sqrtm still pass
• Added backward gradcheck for TorchBackend.sqrtm in test/test_backend.py

PR checklist

• I have read the CONTRIBUTING document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR and Issue fix to the RELEASES.md file.