Fall back to CPU for RoundCeil and RoundFloor expressions

[nartal1]

This contributes to test failures in #5769 esp test_floor* and test_ceil* tests.
While debugging above failures found another bug where a simple ceil on Decimal type was failing.

scala> df.printSchema
root
 |-- value: decimal(18,1) (nullable = true)


scala> df.selectExpr("ceil(value)").collect
22/06/08 17:31:32 WARN GpuOverrides: 
*Exec <ProjectExec> will run on GPU
  *Expression <Alias> CEIL(cast(value#35 as decimal(18,1))) AS CEIL(value)#45 will run on GPU
    *Expression <Ceil> CEIL(cast(value#35 as decimal(18,1))) will run on GPU
      *Expression <Cast> cast(value#35 as decimal(18,1)) will run on GPU
  !Exec <FileSourceScanExec> cannot run on GPU because CSV reading is not 100% compatible when reading doubles. To enable it please set spark.rapids.sql.csv.read.double.enabled to true.

res10: Array[org.apache.spark.sql.Row] = Array([68719476737]

Looks like it was introduced while working on RoundCeil where the maxPrecision was added. Above bug is fixed when the maxPrecision calculation is reverted.
Currently we support RoundCeil/RoundFloor only when scale=0. We have a case where the tests fail when DecimalType(18,0) is given as input. So we are falling back to CPU whenever there is RoundCeil and RoundFloor expr in this PR.
Created a follow-on issue to debug above case and enable these expressions - #5797

[nartal1]

build