Drop support of 8-bit integers input for CumulativeSum

[huningxin]

None backend of Chromium prototype supports 8-bit integers for CumulativeSum operator:

• CoreML cumsum: fp16, fp32, i32
• DirectML DML_CUMULATIVE_SUMMATION_OPERATOR_DESC: FLOAT32, FLOAT16, INT64, INT32, UINT64, UINT32
• ONNX CumSum: tensor(bfloat16), tensor(double), tensor(float), tensor(float16), tensor(int32), tensor(int64), tensor(uint32), tensor(uint64)
• TFLite cumsum: tensor of 32-bit float or 32-bit signless integer or 64-bit signless integer values

For spec-wise consistency, similar reduceSum operator doesn't support 8-bit integers either.

The proposal is to remove the support of 8-int integers input from CumulativeSum tensor limits

operand	allowed data types	allowed ranks
input	"float32", "float16", "int32", "uint32", "int64", "uint64"	1 to N
output	same as input	same as input

[fdwr]

Fine with me (I originally only included float32/float64: 3e0be03 in #805).

Also 8-bit cumulative summation isn't that useful anyway because 8 bits easily overflows.