fix: reject a provider named like an attribute of the target module in expose()

[lesnik512]

Closes #68

expose() installed each fixture with setattr, so a provider whose name was already an attribute of the target module replaced it without a message. A conftest that defined its own settings fixture and then called expose(Deps) with a settings provider lost its fixture.

expose() now raises ValueError when a provider name is already defined on the target module, whether that attribute is a fixture, an imported name or anything else. The message names the attribute and the group, says the module already defines it, and points to renaming or modern_di_fixture(..., name=...). The check runs before anything is installed, like the existing collision checks.

Design notes:

• Per ADR-0001, _collect_fixtures stays pure. It takes a new existing_names argument, and expose() passes vars(module).keys(). To do that, expose() now resolves the target module before collecting. An empty call still raises TypeError first: expose() checks for groups before it looks up the module.
• The container fixture check from A provider named di_container replaces the container fixture and makes every exposed fixture recursive #67 runs first, so a provider named like a container fixture defined in the same module still gets that more specific message.
• Calling expose() twice with the same group on one module now raises, since the second call finds the names from the first.

README: a new "Fixture names and overrides" section says exposed fixtures follow pytest's override rules for fixtures from pytest, plugins and parent conftests, how to avoid a clash, and that expose() should come after the conftest's own fixtures because a later definition cannot be detected.

Out of scope, per the issue: detecting pytest built-in, plugin or parent conftest fixtures, a pytest11 entry point, and rename or prefix options for expose().