Copy error cases from reference tokenizer

[c42f]

The tokenizer is too permissive in some cases, for example some numeric literals

We should survey the reference tokenizer code and make sure all the existing error cases are covered in the new tokenizer. Known problems:

• Bad numeric literals such as 0.0.0, as in Perform error checking for more literals, e.g. 0.0.0 Tokenize.jl#111
• Bad error message for 1.+1 #57
• missing error for ** #112
• a <--- b
• a -- b
• 1._2
• 0b12

[c42f]

Fixed by #196