Pass through non-UTF8 bytes in lines preprocessor

[al2o3cr]

Attempting to CSV.decode a stream that contains non-UTF8 bytes raises a FunctionClauseError:

     ** (FunctionClauseError) no function clause matching in CSV.Decoding.Preprocessing.Lines.starts_sequence?/5

     The following arguments were given to CSV.Decoding.Preprocessing.Lines.starts_sequence?/5:

         # 1
         <<225, 110, 100, 101, 122>>

         # 2
         "n"

         # 3
         false

         # 4
         44

         # 5
         ""

     Attempted function clauses (showing 5 out of 5):

         defp starts_sequence?(<<34::utf8(), tail::binary()>>, last_token, false, separator, _) when last_token == <<separator::utf8()>>
         defp starts_sequence?(<<34::utf8(), tail::binary()>>, "", false, separator, _)
         defp starts_sequence?(<<34::utf8(), tail::binary()>>, _, quoted, separator, sequence_start)
         defp starts_sequence?(<<head::utf8(), tail::binary()>>, _, quoted, separator, sequence_start)
         defp starts_sequence?("", _, quoted, _, sequence_start)

     code: result = CSV.decode(stream) |> Enum.to_list()
     stacktrace:
       (csv 2.4.1) CSV.Decoding.Preprocessing.Lines.starts_sequence?/5
       (csv 2.4.1) lib/csv/decoding/preprocessing/lines.ex:85: CSV.Decoding.Preprocessing.Lines.start_sequence/3
       (elixir 1.13.0) lib/stream.ex:902: Stream.do_transform_user/6

This makes it impossible to handle encoding errors per-line or use machinery like Decoder's replacement option.

The code that would prevent this crash was accidentally deleted in 4f5069b because it is "unused" for files that only contain valid UTF8.

This PR restores the deleted clause and adds a high-level test; existing tests cover Decoder and Lexer but not the complete pipeline.

[aezell]

We are running into this same error with similar data.

[nwai90]

Same here, experienced the issue with non UTF-8 characters