Add sentinel Eof token at the end of the tokenized buffer. (#388)

This has two goals:

1) It allows us to simplify and remove special cases from the parser:
   when we expect a particular token next, we can just check for it
   without needing a special case for end-of-file.

2) It gives us a token to use as a position when emitting diagnostics at
   the end of the file.

Centralize all updating of `position` to `Consume` and `SkipTo`, so that we can in a single place ensure that we never go past the EOF token.

lexer/tokenized_buffer.cpp

@@ -463,6 +463,12 @@ class TokenizedBuffer::Lexer {
     buffer.has_errors = true;
     return token;
   }
+
+  auto AddEndOfFileToken() -> void {
+    buffer.AddToken({.kind = TokenKind::EndOfFile(),
+                     .token_line = current_line,
+                     .column = current_column});
+  }
 };
 
 auto TokenizedBuffer::Lex(SourceBuffer& source, DiagnosticConsumer& consumer)
@@ -491,6 +497,7 @@ auto TokenizedBuffer::Lex(SourceBuffer& source, DiagnosticConsumer& consumer)
   }
 
   lexer.CloseInvalidOpenGroups(TokenKind::Error());
+  lexer.AddEndOfFileToken();
   return buffer;
 }
 
@@ -546,6 +553,10 @@ auto TokenizedBuffer::GetTokenText(Token token) const -> llvm::StringRef {
     return relexed_token->Text();
   }
 
+  if (token_info.kind == TokenKind::EndOfFile()) {
+    return llvm::StringRef();
+  }
+
   assert(token_info.kind == TokenKind::Identifier() &&
          "Only identifiers have stored text!");
   return GetIdentifierText(token_info.id);