Fix HardShrink to accept its documented lambd argument

[Pablosinyores]

What

nn.HardShrink documents a configurable lambd argument (Default: 0.5), but the class cannot actually be constructed with one:

nn.HardShrink(lambd=0.3)
# TypeError: Module.__init__() got an unexpected keyword argument 'lambd'
nn.HardShrink(0.3)
# TypeError: ... takes 1 positional argument but 2 were given

The class is built with @_make_activation_module(hard_shrink), which sets __call__ to always apply the default lambd and leaves the class with no __init__ (it inherits Module.__init__(self)). So the documented parameter is silently unusable.

Fix

Replace the decorator with an explicit __init__/__call__ that stores and forwards lambd, mirroring the sibling Softshrink layer (which already does this and is tested). Every other parameterized activation (Softshrink, ELU, CELU, LeakyReLU, PReLU, Step) follows this same pattern; HardShrink was the lone one mistakenly decorated.

Tests

Extended test_hard_shrink to exercise the class path (nn.HardShrink() and nn.HardShrink(lambd=0.1)), which was previously untested — only the functional nn.hard_shrink was covered, which is why the missing parameter went unnoticed.

[zcbenz]

Thanks!