Fix crash in pylsp_definitions due to definitions without a source position

[metheoryt]

Problem

pylsp_definitions raises on some textDocument/definition requests:

TypeError: unsupported operand type(s) for -: 'NoneType' and 'int'
  File ".../pylsp/plugins/definition.py", line 69, in pylsp_definitions
    "start": {"line": d.line - 1, "character": d.column},

Reported in #623 and #624.

Root cause

The default follow_builtin_definitions is True, and the return-list filter is:

for d in definitions
if d.is_definition() and (follow_builtin_defns or _not_internal_definition(d))

_not_internal_definition(d) is the only place d.line/d.column are checked for
None, but the follow_builtin_defns or ... short-circuit means it is never
evaluated when the setting is left at its default. When jedi resolves a definition
into a compiled/builtin module, d.line and d.column are None, so d.line - 1
crashes.

Fix

Exclude definitions that have no line/column unconditionally — such a definition
cannot be expressed as a valid LSP range regardless of the
follow_builtin_definitions setting. This is a one-clause addition to the existing
filter and leaves the behaviour of builtins that do carry a position (e.g. the
builtins.pyi stub covered by test_builtin_definition) unchanged.

Test

Adds test_definition_without_source_position, which patches jedi's goto to
return a definition with line=None/column=None and asserts pylsp_definitions
returns [] instead of raising. It fails on develop (with the exact TypeError
above) and passes with this change.

Design note

I went with dropping positionless definitions since they can't form a usable LSP
location. An alternative would be to keep them and clamp to line 0 ((d.line or 1) - 1)
so navigation still lands at the top of module_path when one exists — happy to
switch to that if maintainers prefer preserving the result over dropping it.

Fixes #623

[ccordoba12]

Looks good to me, thanks for your contribution @metheoryt!