Default casting to int8 leads to easy overflow errors

Describe the issue:

Minimal example below but if you pass an 'int8' to the gamma distribution, you cannot draw from it. If you make it int16, however, no problem. This problem can show up unexpectedly when a chain of pytensor operations downcasts.

It is also sensitive to the value of mu. If mu is 51, it doesn't happen.

Reproduceable code example:

import pymc as pm
import pytensor.tensor as pt

mu = pt.as_tensor(50,dtype="int8")

rv = pm.Gamma.dist(
    mu = mu,
    sigma = 1
)
pm.draw(rv)

Error message:

ValueError: shape < 0
Apply node that caused the error: gamma_rv{0, (0, 0), floatX, True}(RandomGeneratorSharedVariable(<Generator(PCG64) at 0x1C107CE2260>), [], 11, -60.0, 0.02)
Toposort index: 0
Inputs types: [RandomGeneratorType, TensorType(int64, shape=(0,)), TensorType(int64, shape=()), TensorType(float64, shape=()), TensorType(float64, shape=())]
Inputs shapes: ['No shapes', (0,), (), (), ()]
Inputs strides: ['No strides', (0,), (), (), ()]
Inputs values: [Generator(PCG64) at 0x1C107CE2260, array([], dtype=int64), array(11, dtype=int64), array(-60.), array(0.02)]
Outputs clients: [['output'], ['output']]

PyMC version information:

pymc: 5.14.0
pytensor: 2.20.0

Context for the issue:

No response

Start by running the minimal PyMC and PyTensor example and trace the gamma_rv inputs shown in the error, especially the int8-derived values and resulting negative shape. The issue is done when an int8 mu no longer causes overflow or a negative shape during Gamma sampling, with regression coverage for the reported example.