Fix typos under aten directory (pytorch#87754)

This PR fixes typos in `.md` files under aten directory

Pull Request resolved: pytorch#87754
Approved by: https://github.com/kit1980

aten/src/ATen/native/README.md

@@ -445,7 +445,7 @@ By default, ATen code generation will generate device check,
 which will ensure all the tensor parameters passed to kernel are
 on the same device.
 
-However, in some cases, checking the device is unncessary, because,
+However, in some cases, checking the device is unnecessary, because,
 e.g., you call a function allows to work on multiple devices.
 In that case, code generation of the device check can be disabled by adding
 `device_check: NoCheck` to your function definition.
@@ -556,7 +556,7 @@ Here're steps to follow to decide the right dispatch keyword:
       Note: to support training, you're required to write a formula in
       derivatives.yaml since your backend implementations don't support autograd.
 
-    - Yes: you're likely calling other `at::` ops in the implemetation. Go to step 2.
+    - Yes: you're likely calling other `at::` ops in the implementation. Go to step 2.
 
 2. Think about training: does your kernel support autograd? [check autograd support](#will-your-function-be-automatically-differentiable)
     - Yes: in other words, you're providing a `CompositeImplicitAutograd` kernel which supports both inference and autograd.
@@ -610,7 +610,7 @@ It shows for a certain operator, what the computed dispatch table looks like aft
 4. TODO: AutogradCPUOrCUDA
 
 Note that in native_functions.yaml you can mix using backend keywords and alias keywords above for one op:
-  - direct registration to backend always has higher precendence than alias
+  - direct registration to backend always has higher precedence than alias
   - DO NOT provide multiple alias keywords to the same op: alias keywords have precedence `CompositeExplicitAutograd > CompositeImplicitAutograd`,
     e.g. adding both `CompositeImplicitAutograd` and `CompositeExplicitAutograd` kernels for one op will completely ignore `CompositeImplicitAutograd` kernel for
     both inference and training. Thus this will trigger an error when native_functions.yaml is parsed.

aten/src/ATen/native/cpu/README.md

@@ -64,7 +64,7 @@ within 256bit & 512bits registers. vec defines various operators such as
 As an example `ReduceOpsKernel.cpp` implements a generic `kernel_` that reduces
 an entire array using a given associative binary operation such as +.
 
-More explicity, calling `kernel_` with template argument `std::plus` will cause
+More explicitly, calling `kernel_` with template argument `std::plus` will cause
 it to sum up the entire array into a single value.
 
 `ReduceOpsKernel.cpp` uses the `CPU_CAPABILITY_*` macros to "know" under which
@@ -73,7 +73,7 @@ generic code, which will be compiled under multipled compilation settings.
 
 `../ReduceOps.cpp` now includes the header `ReduceOpsKernel.h`, which contains
 a generic definition of `sumImplAll`. This function allows the user to reduce
-over a dimension or all dimensions. The appropiate capability is chosen at
+over a dimension or all dimensions. The appropriate capability is chosen at
 runtime using cpuinfo. If the current platform has AVX2, `sumImpl` will be set
 to `sumImplAll<CPUCapability::AVX2>`.