Fix layouts in flambda1 partial application simplification

[ccasin]

(I'm not confident this is correct, and if it is other problems remain.)

This addresses an flambda1-specific failure I'm seeing over in #1528, which adds unboxed floats to the front-end. In that PR, this program trips an assert in Inline_and_simplify.simplify_partial_application:

let id_float (x : float#) = x

let app f (x : float#) = f x

let id_float' = app id_float

The assert tripped is:

    assert(Lambda.compatible_layout function_decl.A.return_layout result_layout);

This assert is surprising to me. As far as I can tell, result_layout is the result of the application. Since we're looking at a partial application, the result type is a function and the result layout will always be value. On the other hand, the return_layout of the function being partially applied may not be value.

This PR attempts to fix this by dropping the assert and correcting a layout in the resulting function. I can't test it here since there are no other layouts allowed by the front-end. I have tested it in #1528, and it seems to work. But I'm also seeing other flambda1 problems that I'm less sure how to fix, so I don't feel confident - I'll write a slack message about those.

Edit: The things that worried me in the original message did turn out to be a distinct issue, now addressed in #1557

[Ekdohibs]

We might want to replace the existing assert with something like assert (Lambda.compatible_layout Lambda.layout_function result_layout) as well, however.

[ccasin]

We might want to replace the existing assert with something like assert (Lambda.compatible_layout Lambda.layout_function result_layout) as well, however.

Reasonable. I added back the assert, but moved it just before the call to this function, where I think it makes the most sense.

[goldfirere]

Merging with @Ekdohibs's approval and in @mshinwell's absence.

[goldfirere]

Merging with @Ekdohibs's approval and in @mshinwell's absence.