fix bug where AD is not always seeing custom [differentiable] attrs

[marcrasi]

When SILGen generates SIL for an expression like 3 * x it creates a SILFunciton for Float.* but does not populate that SILFunction with data like the list of [differentiable] attrs. The AD pass does not load this data until after running lookUpMinimalDifferentiationTask, so lookUpMinimalDifferentiationTask does not find the custom adjoint. So AD differentiates Float.* wrt the second parameter instead of using the custom adjoint.

The struct_extract differentiation patch that I am working on exposes this, so I split this fix out into its own PR.

[marcrasi]

@swift-ci please test tensorflow

[rxwei]

Do you know why SILGen did not populate the SILFunction with differentiable attrs? It does add attributes like specialize to the function declaration. See https://forums.swift.org/t/help-needed-with-differentiable-attribute-serialization/19454 for context.

loadFunction deserializes the function body, which we should not do in the long term.