Fix 1431

[jonludlam]

The fix for #930 introduced a problem where a local module substitution fails to resolve, which
led to an unrecoverable exception being raised. This commit detects the problem and essentially
leaves the signature as it was found. This means that the fix for #930 is partial (and
always was). The behaviour that this commit introduces is better now, but not perfect.

The original #930 problem is due to a clash of identifier caused by an include with a module
substitution. The fix was to turn the module substitution into a sig ... end with module Foo :=
expression, calculate the result and use that as the body of the signature, hence no clash.
This was applied universally, so any time an include had a substitution in it, we used this code.

However, if the right-hand side of the module substitution was a local module, as in the test in
this PR, odoc cannot resolve this, and hence raises an error. This commit detects that problem
and reverts the signature to its original form.

Thus the new behaviour means that there's still potential for #930, but now the condition for
a problem is that there's a module substitution being included into a signature that has a module
with the same name (the #930 problem), and that the substitution was for a local module (the
cause of #1431)

[panglesd]

Even though this is not a full fix, it is at least strictly better than what we had before, in terms of behaviour.

Slight summary of the description above, since I initially misunderstood it.
We had issue #930. Then, #1400 fixed it, but introduced the other issue, #1431.
This PR restores the pre-#1400 behavior, only whenever #1431 happens. This way, we only have a problem in the "#930 + #1431 at the same time" situation, which is hopefully quite rare.

[jonludlam]

gah I was going to say this was 'no changelog' but I realised there is a need as 3.2.0 is affected by this.