"successful preparation means successful execution" guarantee neglects bad operation inputs

Under certain circumstances, we guarantee that a prepared model can be
executed successfully.  In describing those circumstances, we
neglected to specify that operation input operands must have legal
values for the guarantee to hold.  For example, the guarantee doesn't
hold if an ADD operation has an activation input that is not one of
the defined values; or if a RESHAPE operation has a shape input in
which two or more components are -1.

This change modifies the guarantee to apply only when operation input
operands have legal values.  It also documents this guarantee for
burst execution.

Note that if an operation has an input operand that can be proven to
have an illegal value at preparation time (e.g., a constant value that
is illegal), model preparation might (but is not required to) fail for
that reason.

Bug: 135933040

Test: $ cd neuralnetworks ; mma
Change-Id: I8b421550dd89e4bbbdae899e7cb5e9e88a46d2fb
(cherry picked from commit 48544cc3)