The reason why decryption needs to be poker face is because it is possible to distinguish parts about the key or plaintext after modular exponentiation, when performing the unpadding. The unpadding part is susceptible to padding oracle attacks after all.
Modular operations will map the integer to an arbitrary value within the range [0, N) by simply performing mod N on the input value. That usually the case as this would be the mathematical equivalent - beware that optimized libraries may perform different calculations.
An attacker that is able to change the input message is commonly capable of choosing the equivalent message in the range [0, N) by themselves. As such, the processing of larger or smaller integers shouldn't make any difference to the security of the RSA algorithm.
It would still be a bad idea to remove these checks though:
- they could be used to try and find a buffer overruns or trigger some other kind of overflow;
- the size may have been increased to shift other values of the message within the protocol;
- they might be used to slow down the system as you'd still have the initial modular operation on a number of arbitrary size.
In general: the higher the confidence that the state of the system is within bounds, the higher the security. Pre-conditions and guard statements are relatively cheap ways of making sure that the input is valid and that calculations are not performed using out-of-range values.
