This PR fixes a potential compilation issue that appears in certain edge cases where the explicit typecast during SQL generation somehow confuses the type inference.

This does not seem to be a frequent issue as it only ever happened to me in one occasion where I had two models without date / timestamp columns (the following are abstractions of those)

class Sample < Jennifer::Model::Base
  table_name "samples"

  mapping({
    id:             Int64,
    name:       String?,
    secret:      String?
  })
  

  has_and_belongs_to_many :others, Other, join_table: "sample_others", association_foreign: "other_id", foreign: "sample_id"
end

class Other < Jennifer::Model::Base
  table_name "others"

  mapping({
    id:             Int64,
    title:       String?
  })
end

which are connected via a many-to-many relation (with one-sided access).

When calling sample.remove_others(other), this leads to a corresponding compilation error as the type inference algorithm seems to believe that the cast in

The error is then as follows:

In lib/jennifer/src/jennifer/adapter/postgres/sql_generator.cr:126:15

 126 | args[i] = arg.as(Time).to_utc if arg.is_a?(Time)
           ^
Error: no overload matches 'Array(Bool | Float64 | Int64 | String | Nil)#[]=' with types Int32, Time

Overloads are:
 - Array(T)#[]=(index : Int, count : Int, values : Array(T))
 - Array(T)#[]=(index : Int, count : Int, value : T)
 - Array(T)#[]=(index : Int, value : T)
 - Array(T)#[]=(range : Range, values : Array(T))
 - Array(T)#[]=(range : Range, value : T)

Appendix

Since this seems to be also a bug in the compiler / type inference algorithm, it would probably worth it to also post it there. I wanted to push this here first, as it is easy to fix for this simple case - please let me know what you think about that.