[AutoDiff] Eliminate flow-sensitivity in '@differentiable' closure conversions.

[rxwei]

@differentiable function conversion from a non-@differentiable function is handled by inserting an AutoDiffFunctionExpr in CSApply when the contextual type is @differentiable. This is similar to C function poster conversion. One thing that we have not checked during these conversions is whether the original function expression is a literal closure expression or a reference to a func declaration. Lacking this check made @differentiable closure conversion flow-sensitive, which caused many semantic inconsistencies as described in TF-577.

This patch makes @differentiable closure conversion only valid when the from-expression is a literal closure expression or a reference to a func declaration. This eliminates all flow-sensitivity in @differentiable closure conversions.

In common cases where a user writes a higher-order function that differentiates the argument closure but forgets to make the parameter have a @differentiable or @differentiable(linear) function type, we produce a note and a fix-it to insert a @differentiable or @differentiable(linear) attribute to the parameter's type signature.

[rxwei]

@lattner you wanted this done a few months ago :)

[rxwei]

@swift-ci please test tensorflow

[dan-zheng]

Nice step towards semantic consistency!

[lattner]

Awesome, this is a great step forward!