Free pointers in parser activations (#4486)

* Free pointers in ActivationsC

* Restructure alloc/free for parser activations

* Rewrite/restructure to have allocation and free in parallel functions
in `_parser_model` rather than partially in `_parseC()` in `Parser`.

* Remove `resize_activations` from `_parser_model.pxd`.

spacy/syntax/_parser_model.pxd

@@ -36,7 +36,9 @@ cdef WeightsC get_c_weights(model) except *
 
 cdef SizesC get_c_sizes(model, int batch_size) except *
 
-cdef void resize_activations(ActivationsC* A, SizesC n) nogil
+cdef ActivationsC alloc_activations(SizesC n) nogil
+
+cdef void free_activations(const ActivationsC* A) nogil
 
 cdef void predict_states(ActivationsC* A, StateC** states,
         const WeightsC* W, SizesC n) nogil

spacy/syntax/_parser_model.pyx

@@ -62,6 +62,21 @@ cdef SizesC get_c_sizes(model, int batch_size) except *:
     return output
 
 
+cdef ActivationsC alloc_activations(SizesC n) nogil:
+    cdef ActivationsC A
+    memset(&A, 0, sizeof(A))
+    resize_activations(&A, n)
+    return A
+
+
+cdef void free_activations(const ActivationsC* A) nogil:
+    free(A.token_ids)
+    free(A.scores)
+    free(A.unmaxed)
+    free(A.hiddens)
+    free(A.is_valid)
+
+
 cdef void resize_activations(ActivationsC* A, SizesC n) nogil:
     if n.states <= A._max_size:
         A._curr_size = n.states

spacy/syntax/nn_parser.pyx

@@ -27,7 +27,8 @@ from thinc.neural.util import get_array_module
 from thinc.linalg cimport Vec, VecVec
 import srsly
 
-from ._parser_model cimport resize_activations, predict_states, arg_max_if_valid
+from ._parser_model cimport alloc_activations, free_activations
+from ._parser_model cimport predict_states, arg_max_if_valid
 from ._parser_model cimport WeightsC, ActivationsC, SizesC, cpu_log_loss
 from ._parser_model cimport get_c_weights, get_c_sizes
 from ._parser_model import ParserModel
@@ -312,8 +313,7 @@ cdef class Parser:
             WeightsC weights, SizesC sizes) nogil:
         cdef int i, j
         cdef vector[StateC*] unfinished
-        cdef ActivationsC activations
-        memset(&activations, 0, sizeof(activations))
+        cdef ActivationsC activations = alloc_activations(sizes)
         while sizes.states >= 1:
             predict_states(&activations,
                 states, &weights, sizes)
@@ -327,6 +327,7 @@ cdef class Parser:
                 states[i] = unfinished[i]
             sizes.states = unfinished.size()
             unfinished.clear()
+        free_activations(&activations)
 
     def set_annotations(self, docs, states_or_beams, tensors=None):
         cdef StateClass state