[Bug] [TIR] float infinity not supported in TIR with CUDA target

[shingjan]

isinf is an operator in relay/topi/te. But when we lower relay functions containing isinf to cuda, the generated float infinity is not defined and supported, therefore an compilation error on nvcc is thrown.

Expected behavior

Infinity should be supported on TIR level as there is a function isinf in relay/topi/te.

Actual behavior

Compilation error:
/tmp/tmpr1n6rqid/my_kernel.cu(331): error: identifier "inff" is undefined

/tmp/tmpr1n6rqid/my_kernel.cu(276): warning #177-D: function "__pack_half2" was declared but never referenced

/tmp/tmpr1n6rqid/my_kernel.cu(301): warning #177-D: function "hpow" was declared but never referenced

/tmp/tmpr1n6rqid/my_kernel.cu(302): warning #177-D: function "htanh" was declared but never referenced

/tmp/tmpr1n6rqid/my_kernel.cu(303): warning #177-D: function "htan" was declared but never referenced

/tmp/tmpr1n6rqid/my_kernel.cu(304): warning #177-D: function "hatan" was declared but never referenced

/tmp/tmpr1n6rqid/my_kernel.cu(305): warning #177-D: function "herf" was declared but never referenced

1 error detected in the compilation of "/tmp/tmpr1n6rqid/my_kernel.cu".

Environment

Latest TVM main on Nvidia RTX3070

Steps to reproduce

Minimal repro script can be found here.

cc @Hzfengsy @junrushao @junrushao1994 @zxybazh

[wrongtest-intellif]

"inff" may indicate that the "inf" literal of float32 get mistakenly generated as a normal value like 0.3f.