Use Markdown section marker `#`; ref #5084 [ci skip]

The fix was proposed by @dearblue

doc/mruby3.md

 User visible changes in `mruby3`
 ===
 
-= Build System
+# Build System
 
 You can specify `TARGET` option to `rake` via a command line
 option, or via an environment variable, e.g.
@@ -15,13 +15,13 @@ or
 It's much easier to switch multiple targets than the
 previous `build_config.rb` system.
 
-== `presym` target
+## `presym` target
 
 The first compilation of `mruby` may require generation of a
 static symbol table named `build/presym`. You can generate
 the table by `rake gensym`.
 
-== `target` directory
+## `target` directory
 
 Build target specification files are loaded from `target`
 directory. The default `TARGET` is `host` which is described
@@ -39,15 +39,15 @@ If you want to have your target description out of the
 source tree, you can specify the path to the description
 file in `MRUBY_CONFIG`.
 
-= Build Target Contribution
+# Build Target Contribution
 
 When you write a new target description, please
 contribute. We welcome your contribution as a GitHub
 pull-request.
 
-= Languge Changes
+# Languge Changes
 
-== New Syntax
+## New Syntax
 
 We have ported some new syntax from CRuby.
 
@@ -55,54 +55,54 @@ We have ported some new syntax from CRuby.
 * Numbered block parameter (`x.map{_1 * 2}`)
 * End-less `def` (`def double(x) = x*2`)
 
-= Configuration Options Changed
+# Configuration Options Changed
 
 Some configuration macro names are changed for consistency
 
-== `MRB_NO_FLOAT`
+## `MRB_NO_FLOAT`
 
 Changed from `MRB_WITHOUT_FLOAT` to conform `USE_XXX` naming
 convention.
 
-== `MRB_USE_FLOAT32`
+## `MRB_USE_FLOAT32`
 
 Changed from `MRB_USE_FLOAT` to make sure `float` here means
 using single precision float, and not the opposite of
 `MRB_NO_FLOAT`.
 
-== `MRB_USE_METHOD_T_STRUCT`
+## `MRB_USE_METHOD_T_STRUCT`
 
 Changed from `MRB_METHOD_T_STRUCT`.
 
 To use `struct` version of `mrb_method_t`. More portable but consumes more memory.
 Turned on by default on 32bit platforms.
 
-== `MRB_NO_BOXING`
+## `MRB_NO_BOXING`
 
 Uses `struct` to represent `mrb_value`. Consumes more memory
 but easier to inveticate the internal and to debug. It used
 to be default `mrb_value` representation. Now the default is
 `MRB_WORD_BOXING`.
 
-== `MRB_WORD_BOXING`
+## `MRB_WORD_BOXING`
 
 Pack `mrb_value` in an `intptr_t` integer. Consumes less
 memory compared to `MRB_NO_BOXING` especially on 32 bit
 platforms. `Fixnum` size is 31 bits so some integer values
 does not fit in `Fixnum` integers.
 
-== `MRB_NAN_BOXING`
+## `MRB_NAN_BOXING`
 
 Pack `mrb_value` in a floating pointer number. Nothing
 changed from previous versions.
 
-== `MRB_USE_MALLOC_TRIM`
+## `MRB_USE_MALLOC_TRIM`
 
 Call `malloc_trim(0)` from mrb_full_gc() if this macro is defined.
 If you are using glibc malloc, this macro could reduce memory consumption.
 
-= Internal Changes
+# Internal Changes
 
-== `Random` now use `xoshiro128++`.
+## `Random` now use `xoshiro128++`.
 
 For better and faster random number generation.