Fall back to op.decompose if op is unsolved in decomposition graph

[astralcai]

Context:
The DecompositionGraph currently errors out completely if any of the operators is unsolved for, and as a result, with graph enabled, if a circuit contains a single operator/template that is not yet integrated with the new decomposition system, the entire thing fails, and the user is forced to turn off graph mode and use the old decomposition system. We want to change this behaviour, make it so that the DecompositionGraph raise a warning instead of an error if certain operators are unsolved for, and fall back to using op.decomposition for those operators.

Description of the Change:

• Downgrade error to a warning when certain operators are unsolved in the graph.
• Add sensible warning for when the graph failed to find a solution due to GlobalPhase.
• When the new system is enabled, operators without a decomposition will be left undecomposed with a warning.

Benefits:

• The user can use the new system for the rest of the circuit even if it contains templates that are not yet integrated with the new graph-based system.

Possible Drawbacks:

Related GitHub Issues:
[sc-98390]

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 99.69%. Comparing base (8612788) to head (d8d78fb).
⚠️ Report is 3 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8156      +/-   ##
==========================================
- Coverage   99.69%   99.69%   -0.01%     
==========================================
  Files         548      548              
  Lines       56815    56805      -10     
==========================================
- Hits        56641    56631      -10     
  Misses        174      174              

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[JerryChen97]

Thanks for implementing this fallback logic! Would you like to add a test over a dummy operator with no matrix to see it gets channelled back to old decomposition and emits DeviceError as well? Something like:

class NoMatNoDecompOp(qml.operation.Operation):
    """Dummy operation for checking check_validity throws error when
    expected."""

    # pylint: disable=arguments-renamed, invalid-overridden-method
    @property
    def has_matrix(self):
        return False

[astralcai]

Thanks for implementing this fallback logic! Would you like to add a test over a dummy operator with no matrix to see it gets channelled back to old decomposition and emits DeviceError as well? Something like:

class NoMatNoDecompOp(qml.operation.Operation):
    """Dummy operation for checking check_validity throws error when
    expected."""

    # pylint: disable=arguments-renamed, invalid-overridden-method
    @property
    def has_matrix(self):
        return False

That would be irrelevant to this PR. The behaviour you're talking about is exclusive to preprocess.decompose.

[JerryChen97]

Thanks for implementing this fallback logic! Would you like to add a test over a dummy operator with no matrix to see it gets channelled back to old decomposition and emits DeviceError as well? Something like:

class NoMatNoDecompOp(qml.operation.Operation):
    """Dummy operation for checking check_validity throws error when
    expected."""

    # pylint: disable=arguments-renamed, invalid-overridden-method
    @property
    def has_matrix(self):
        return False

That would be irrelevant to this PR. The behaviour you're talking about is exclusive to preprocess.decompose.

This PR does not involve any Operator with has_matrix=False?

[astralcai]

Thanks for implementing this fallback logic! Would you like to add a test over a dummy operator with no matrix to see it gets channelled back to old decomposition and emits DeviceError as well? Something like:

class NoMatNoDecompOp(qml.operation.Operation):
    """Dummy operation for checking check_validity throws error when
    expected."""

    # pylint: disable=arguments-renamed, invalid-overridden-method
    @property
    def has_matrix(self):
        return False

That would be irrelevant to this PR. The behaviour you're talking about is exclusive to preprocess.decompose.

This PR does not involve any Operator with has_matrix=False?

Whether or not an operator has matrix is completely irrelevant here. The has_matrix check is part of DefaultQubit's stopping condition. This PR is about making it so that when the new graph-based system is not able to find a decomposition for an operator, use op.decomposition() as a fall back. This has nothing to do with the stopping condition whatsoever.